Automatic driving decision-making method and system for uncertain scene
By building multiple prediction and decision models in autonomous trucks, accurately quantifying and processing uncertain scenarios, the problem that autonomous trucks cannot maintain sufficient safe distance in uncertain scenarios is solved, and safer and more efficient driving is achieved.
Patent Information
- Application Number
- CN202411950912.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-27
- Publication Date
- 2025-04-18
AI Technical Summary
In the autonomous truck scenario, the prior art is difficult to effectively deal with uncertain scenarios, resulting in the inability to maintain sufficient safe distances and increase safety risks.
By constructing a classified prediction model, probability calibration model, inference prediction model, probability tracking and correction model and decision output model, we accurately quantify the uncertainty of scene interaction and make decisions based on probability to ensure safety and efficiency.
Accurate quantification and safety decision-making of uncertain scenarios are achieved, safety risks are avoided, truck driving efficiency is improved, and it is more in line with the habits and risk expectations of human drivers.
Smart Images
Figure CN119840655A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of autonomous driving information processing technology, and in particular to an autonomous driving decision-making method and system for uncertain scenarios. Background Art
[0002] An autonomous vehicle is an intelligent vehicle that senses the road environment through an on-board sensor system, automatically plans a driving route, and controls the vehicle to reach a predetermined destination. Adaptive cruise control is one of the basic functions of an autonomous vehicle, which allows the vehicle to automatically adjust its speed to adapt to traffic conditions without driver intervention. With the development of adaptive cruise control systems, car owners are using adaptive cruise control systems more and more frequently. In the conventional interactive decision-making method of passenger cars, predictions are usually made for surrounding vehicles first, and decisions are made for the vehicle based on this. In the passenger car scenario, the other vehicle's intention to change lanes can usually be output when the other vehicle has made a lateral move. At this time, the other vehicle's intention is very clear and the uncertainty is small, and the vehicle can slow down according to the situation. When the other vehicle is preparing to change lanes, such as accelerating to overtake the vehicle, the uncertainty of the other vehicle's intention is large, and the vehicle usually does not respond. In addition, due to the characteristics of trucks such as heavy weight, large inertia, and slow response, these characteristics further increase the difficulty of trucks in handling the uncertainty of interactive scenarios. If the decision-making scheme of passenger cars is directly applied to the truck autonomous driving scenario, it is often impossible to maintain a sufficient safety distance, resulting in safety risks. Summary of the invention
[0003] In view of the above problems, the first purpose of the present invention is to provide an automatic driving decision-making method for uncertain scenarios, which can accurately quantify the uncertainty of scene interaction and give decision instructions based on it. It can accurately and fully consider potential risks, making automatic driving trucks safer. At the same time, the precise quantification of uncertainty and the probability-based decision-making method will not overreact even in low-probability situations. While ensuring safety, it improves the driving efficiency of trucks and is more in line with the habits and risk expectations of human drivers.
[0004] A second object of the present invention is to provide an autonomous driving decision-making system for uncertain scenarios.
[0005] To achieve the first objective, the first technical solution of the present invention is: an automatic driving decision-making method for uncertain scenarios, comprising:
[0006] Step S01: construct a classification prediction model to receive a data set, classify and predict the intention and trajectory mode of other vehicles, and output the classification prediction results;
[0007] The data set is collected through road testing, including the intention and trajectory of other vehicles;
[0008] Step S02: constructing a probability calibration model, for mapping and calibrating the classification prediction results output by the classification prediction model with the probability of lane change of another vehicle, and outputting prior information of the probability of lane change of another vehicle;
[0009] Step S03: constructing an inference prediction model, predicting the intention of the other vehicle according to the prior information of the lane change probability of the other vehicle and the real-time behavior information of the other vehicle, and obtaining the real-time intention probability of the other vehicle;
[0010] Step S04: constructing a probability tracking and correction model to correct the real-time intention probability of the other vehicle;
[0011] Step S05: construct a decision output model for making a self-vehicle behavior decision based on the corrected real-time intention probability of the other vehicle.
[0012] Preferably, step S03 further includes: constructing a multi-information source extraction model for obtaining the position, trajectory, and category of the other vehicle as the real-time behavior information of the other vehicle, and outputting the real-time behavior information of the other vehicle to the inference prediction model.
[0013] Preferably, step S04 further includes: screening abnormal behavior information of other vehicles, and outputting the abnormal behavior information of other vehicles to the probability tracking and correction model.
[0014] Preferably, the abnormal behavior information of the other vehicle includes: the other vehicle crossing the line, irregular speed change, irregular lane change, and abnormal change in the distance between the other vehicle and the own vehicle.
[0015] Preferably, the probability tracking and correction model uses a particle filtering method to correct the real-time intention probability of the other vehicle.
[0016] Preferably, the particle filtering method includes: representing the real-time intention probability of the other vehicle as a particle set, initializing the particles therein with the same weight; performing particle filtering to update the particles in the particle set to obtain updated particles, and comparing them with the abnormal behavior information of the other vehicle; and correcting the real-time intention probability of the other vehicle.
[0017] Preferably, comparing the updated particles with the abnormal behavior information of the other vehicle includes:
[0018] If the comparison result shows a high similarity, the weight of the particle is increased;
[0019] If the comparison result is that the similarity is low, the weight of the particle is reduced.
[0020] Preferably, the classification prediction model and probability calibration model are offline models;
[0021] The inference prediction model, probability tracking and correction model, and decision output model are online models.
[0022] To achieve the second objective, the second technical solution of the present invention is: an automatic driving decision system for uncertain scenarios, comprising:
[0023] Offline module: used to receive a data set, classify and predict the intention and trajectory mode of the other vehicle, and output the classification prediction result; the data set is collected through road testing, including the intention and trajectory of the other vehicle; the classification prediction result is mapped and calibrated with the probability of the other vehicle changing lanes, and the prior information of the probability of the other vehicle changing lanes is output;
[0024] Online module: predict the intention of the other vehicle based on the prior information of the lane change probability of the other vehicle and the real-time behavior information of the other vehicle to obtain the real-time intention probability of the other vehicle; correct the real-time intention probability of the other vehicle; and make the behavior decision of the own vehicle based on the corrected real-time intention probability of the other vehicle.
[0025] Preferably, the online module is also used to obtain the location, trajectory, and category of the other vehicle as the real-time behavior information of the other vehicle and to filter the abnormal behavior information of the other vehicle.
[0026] Beneficial effects of the above technical solution:
[0027] The automatic driving decision-making method and system for uncertain scenarios provided by the present invention are applied to heavy / light box / trailer trucks with automatic driving functions. It can handle complex dynamic traffic scenarios, accurately quantify the uncertainty of scene interactions, and give decision instructions based on this. It avoids the difficulty of handling the uncertainty of interactive scenarios due to the characteristics of heavy truck weight, large inertia, slow response, etc. It solves the problem that if the decision-making scheme of passenger cars is directly applied to the automatic driving scenario of trucks, it often cannot maintain a sufficient safe distance and cause safety risks. The method provided by the present invention can accurately and fully consider potential risks, making automatic driving trucks safer. At the same time, the precise quantification of uncertainty and the probability-based decision-making method will not overreact even in low-probability situations. While ensuring safety, it improves the driving efficiency of trucks and is more in line with the habits and risk expectations of human drivers. It can make automatic driving trucks safer, more efficient, and more in line with the habits of human drivers. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0029] Figure 1 A flow chart of an autonomous driving decision-making method for uncertain scenarios provided by the present invention;
[0030] Figure 2 A flowchart of autonomous driving decision-making for uncertain scenarios provided by the present invention. DETAILED DESCRIPTION
[0031] The following is a further detailed description of the implementation methods of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.
[0032] The terms "first", "second", etc. (if any) in the specification and claims are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.
[0033] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.
[0034] Embodiment 1
[0035] In this embodiment, an automatic driving decision-making method for uncertain scenarios is provided, which can quantify the uncertainty in the scenario in the form of probability and make driving decisions based on probability. It mainly consists of two parts: offline and online. Among them, the classification prediction model and the probability calibration model are offline models. The inference prediction model, the probability tracking and correction model, the decision output model and the multi-information source extraction model are online models.
[0036] Classification prediction model: Use large-scale data sets to design and train neural network models, use the perception output results as network input, and perform classification predictions on the vehicle’s intentions, trajectory modes, etc.
[0037] Probabilistic calibration model: Utilize large-scale data sets to correctly map the output of the prediction model into the probability space, providing model prior information for the reasoning of the online model.
[0038] Multi-information source extraction model: Based on the multiple sensors configured in the self-driving truck, multiple information sources are extracted to provide multiple information inputs for prediction model reasoning and online probability correction.
[0039] Inference prediction model: Predicts the intention probability of surrounding vehicles based on the results of the offline part.
[0040] Probability tracking and correction model: Based on the probability value of the inference prediction model, the scene features brought by the multi-information source extraction model are used. Based on the principle of Bayesian filtering, the probability value is tracked and corrected between multiple time frames, thereby solving the problems of inaccurate and unstable probability caused by distribution drift.
[0041] Decision output model: Based on the corrected probability value of other vehicles’ intentions, analyze the risk level and make corresponding decisions.
[0042] The specific process is as follows Figure 1 As shown. Includes:
[0043] An automatic driving decision-making method for uncertain scenarios, characterized by comprising:
[0044] Step S01: constructing a classification prediction model for receiving a data set, classifying and predicting the intention and trajectory modality of another vehicle, and outputting a classification prediction result; the data set is collected through road testing, including the intention and trajectory of another vehicle;
[0045] Step S02: constructing a probability calibration model, for mapping and calibrating the classification prediction results output by the classification prediction model with the probability of lane change of another vehicle, and outputting prior information of the probability of lane change of another vehicle;
[0046] Step S03: constructing an inference prediction model, predicting the intention of another vehicle based on the prior information of the lane change probability of the other vehicle and the real-time behavior information of the other vehicle, and obtaining the real-time intention probability of the other vehicle; constructing a multi-information source extraction model for obtaining the position, trajectory, and category of the other vehicle as the real-time behavior information of the other vehicle, and outputting the real-time behavior information of the other vehicle to the inference prediction model,
[0047] Step S04: Construct a probability tracking and correction model to correct the probability of the real-time intention of the other vehicle; filter the abnormal behavior information of the other vehicle, and output the abnormal behavior information of the other vehicle to the probability tracking and correction model. The abnormal behavior information of the other vehicle includes: the other vehicle crossing the line, irregular speed change, irregular lane change, and abnormal change in the distance between the other vehicle and the vehicle. The probability tracking and correction model uses a particle filtering method to correct the probability of the real-time intention of the other vehicle.
[0048] Step S05: construct a decision output model for making a self-vehicle behavior decision based on the corrected real-time intention probability of the other vehicle.
[0049] The particle filtering method includes: representing the real-time intention probability of other vehicles as a particle set, initializing the same weight for the particles in the particle set; performing particle filtering to update the particles in the particle set to obtain updated particles, and comparing them with the abnormal behavior information of other vehicles output by the multi-information source extraction model; correcting the real-time intention probability of other vehicles. The updated particles are compared with the abnormal behavior information of other vehicles output by the multi-information source extraction model; when the comparison result is highly similar, the weight of the particle is increased; when the comparison result is less similar, the weight of the particle is reduced.
[0050] Embodiment 2
[0051] In this embodiment, an automatic driving decision-making method for an uncertain scenario includes:
[0052] Classification prediction model: Taking the prediction of other vehicles’ intention to change lanes or go straight as an example, we can use real road test data, annotate the perception results and other vehicles’ intention classification as training sets. We can design convolutional networks, recurrent neural networks, attention-based Transformer networks, etc. for training and output the probability value of lane change intention.
[0053] Probability calibration model: The Platt scaling algorithm can be used to fit the Logistic Regression model to the probability values output by the above model. The output of this model is the calibrated probability output, that is, the prior information on the probability of the other vehicle changing lanes.
[0054] Multi-information source extraction model: Existing perception algorithms and modules can be used to provide inputs such as the position, trajectory, and category of other vehicles to the prediction model. Perception outputs can also be used to filter out typical threat scenarios such as vehicles crossing the line, vehicles with a slow vehicle in front of the left side of the vehicle but approaching at high speed from the left rear, and cones according to rules as inputs for probabilistic tracking and correction.
[0055] Inference prediction model: Use the TensorRT tool to deploy offline trained prediction models and probability calibration models to predict the lane change probability of other vehicles as input for probability tracking and correction.
[0056] Probabilistic tracking and correction model: The particle filtering method can be used for probabilistic tracking and correction, which specifically includes: representing the probability distribution / uncertainty of the intention in the form of a particle set, where each particle represents the intention of a certain other car or a combination of intentions of the entire scene. Therefore, a particle set, that is, a bunch of particles, can be used as an approximation of the probability distribution. When the particle filter is updated, it is first assumed that the probability value will not jump between two consecutive frames, so the state transfer equation is defined as adding Bernoulli noise on the basis of unchanged intention. According to the definition of the state equation, each particle predicts and propagates to the next time step. During the propagation process, the particle changes the intention it represents with a certain small probability, and keeps the intention unchanged in other cases. At the same time, consider the probability measurement obtained based on the perception of the current scene. For each particle, this probability measurement can be used to calculate the similarity between the particle and the real scene. If the intention Figure 1 If the update weight is high, it means the similarity is high; otherwise, it means the similarity is low. Based on the similarity, the update weights of all particles can be calculated. If the update weight is high, there are more opportunities to be sampled multiple times during the resampling process, thereby increasing the proportion of the particle in the overall particle set; if the update weight is low, the particle is gradually abandoned. When the particle filter converges, the distribution of the particle set can be used as an approximation of the real scene distribution. Further, consider two types of probability measurements. One type of probability measurement is the output of the calibrated prediction model, which represents the prior knowledge learned from offline data. The other type of probability measurement is the possibility of the other car changing lanes in the threat scenario, which represents the information learned from online direct observation. For example, even if there is data distribution drift and the prediction model has not learned the impact of cones on lane change intentions from offline data, we can also perform probability updates of multiple information sources through direct online recognition of cones. Obviously, if there is a cone blocking the road ahead, the probability that the other car intends to go straight will be greatly reduced.
[0057] Decision output model: Based on good probability output, a typical implementation method can be to define the problem as a partially observable Markov decision process, perform Monte Carlo tree search in the probability space, and minimize the cost function under probability weighting to obtain the optimal decision considering the probability distribution.
[0058] Preferably, in a specific embodiment of the present invention, the prediction model may also adopt other model structures; the probability calibration model may also adopt algorithms such as temperature scaling or isotonic regression; and the probability-based decision may also be processed by rules or neural networks.
[0059] The method provided by the present invention can accurately quantify the uncertainty of scene interaction and give decision instructions based on it. It can accurately and fully consider potential risks, making self-driving trucks safer. At the same time, the accurate quantification of uncertainty and the probability-based decision-making method will not overreact even in low-probability situations. While ensuring safety, it improves the driving efficiency of trucks and is more in line with the habits and risk expectations of human drivers.
[0060] The present invention provides an automatic driving decision-making system for uncertain scenarios, comprising:
[0061] Offline module: used to receive the data set, classify and predict the intention and trajectory mode of the other vehicle, and output the classification prediction result; map and calibrate the classification prediction result with the lane change probability of the other vehicle, and output the prior information of the lane change probability of the other vehicle;
[0062] Online module: predict the intention of the other vehicle based on the prior information of the lane change probability of the other vehicle and the real-time behavior information of the other vehicle to obtain the real-time intention probability of the other vehicle; correct the real-time intention probability of the other vehicle; and make the behavior decision of the own vehicle based on the corrected real-time intention probability of the other vehicle.
[0063] The online module also includes a method for obtaining the position, trajectory, and category of other vehicles as real-time behavior information of other vehicles and filtering abnormal behavior information of other vehicles.
[0064] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made on the basis of the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the protection scope of the present invention.
Claims
1. An autonomous driving decision-making method for uncertain scenarios, characterized in that: include: Step S01: construct a classification prediction model to receive a data set, classify and predict the intention and trajectory mode of other vehicles, and output the classification prediction results; The data set is collected through road testing, including the intention and trajectory of other vehicles; Step S02: constructing a probability calibration model, for mapping and calibrating the classification prediction results output by the classification prediction model with the probability of lane change of another vehicle, and outputting prior information of the probability of lane change of another vehicle; Step S03: constructing an inference prediction model, predicting the intention of the other vehicle according to the prior information of the lane change probability of the other vehicle and the real-time behavior information of the other vehicle, and obtaining the real-time intention probability of the other vehicle; Step S04: constructing a probability tracking and correction model to correct the real-time intention probability of the other vehicle; Step S05: construct a decision output model for making a self-vehicle behavior decision based on the corrected real-time intention probability of the other vehicle.
2. The automatic driving decision-making method for uncertain scenarios according to claim 1, characterized in that: Step S03 also includes: constructing a multi-information source extraction model for obtaining the position, trajectory, and category of the other vehicle as the real-time behavior information of the other vehicle, and outputting the real-time behavior information of the other vehicle to the inference prediction model.
3. The automatic driving decision-making method for uncertain scenarios according to claim 1, characterized in that: Step S04 also includes: screening the abnormal behavior information of other vehicles, and outputting the abnormal behavior information of other vehicles to the probability tracking and correction model.
4. The automatic driving decision-making method for uncertain scenarios according to claim 3, characterized in that: The abnormal behavior information of other vehicles includes: other vehicles crossing the line, irregular speed changes, irregular lane changes, and abnormal changes in the distance between other vehicles and the vehicle.
5. The automatic driving decision-making method for uncertain scenarios according to claim 4, characterized in that: The probability tracking and correction model uses a particle filtering method to correct the real-time intention probability of the other vehicle.
6. The automatic driving decision-making method for uncertain scenarios according to claim 5, characterized in that: The particle filtering method includes: representing the real-time intention probability of the other vehicle by a particle set, initializing the same weights for the particles therein; performing particle filtering update on the particles in the particle set to obtain updated particles, and comparing them with the abnormal behavior information of the other vehicle; and correcting the real-time intention probability of the other vehicle.
7. The automatic driving decision-making method for uncertain scenarios according to claim 6, characterized in that: The comparison between the updated particles and the abnormal behavior information of the other vehicle includes: If the comparison result shows a high similarity, the weight of the particle is increased; If the comparison result is that the similarity is low, the weight of the particle is reduced.
8. The automatic driving decision-making method for uncertain scenarios according to claim 1, characterized in that: The classification prediction model and probability calibration model are offline models; The inference prediction model, probability tracking and correction model, and decision output model are online models.
9. An autonomous driving decision system for uncertain scenarios, characterized in that: include: Offline module: used to receive a data set, classify and predict the intention and trajectory mode of the other vehicle, and output the classification prediction result; map and calibrate the classification prediction result with the probability of the other vehicle changing lanes, and output the prior information of the probability of the other vehicle changing lanes; the data set is collected through road testing, including the intention and trajectory of the other vehicle; Online module: predict the intention of the other vehicle based on the prior information of the lane change probability of the other vehicle and the real-time behavior information of the other vehicle to obtain the real-time intention probability of the other vehicle; correct the real-time intention probability of the other vehicle; and make the behavior decision of the own vehicle based on the corrected real-time intention probability of the other vehicle.
10. The automatic driving decision system for uncertain scenarios according to claim 9, characterized in that: The online module is also used to obtain the location, trajectory, and category of the other vehicle as the real-time behavior information of the other vehicle and to filter the abnormal behavior information of the other vehicle.