An autonomous driving car-following control method based on brain-like continuous learning

Abstract

The application discloses an autonomous driving car-following control method based on brain-like continuous learning. The method comprises the following steps: obtaining a driving data set under an incremental task; inputting the driving data set under the incremental task into a vehicle car-following control model trained under a previous task for continuous learning, so as to obtain a vehicle car-following control model trained under the incremental task; collecting driving data of a car-following vehicle in real time, inputting the driving data into the vehicle car-following control model trained under the incremental task, obtaining real-time predicted driving data, and feeding back the real-time predicted driving data to the car-following vehicle; the car-following vehicle adjusts a car-following behavior according to the predicted driving data; collecting driving data of the car-following vehicle after the car-following behavior is adjusted in real time, and taking the collected driving data as driving data of a next task. The application enhances the generalization ability and stability of an automatic driving system, significantly improves the adaptability of a vehicle car-following control system in a dynamic traffic environment, and provides a safer and more efficient autonomous driving solution.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent transportation, and particularly relates to an autonomous driving car-following control method based on brain-like continuous learning. BACKGROUND

[0002] With the rapid development of technology, autonomous driving technology has made significant progress. However, despite this, autonomous driving systems still face many challenges in dealing with complex and unpredictable driving environments. In particular, the "long tail problem", which means that some rare but critical driving scenarios may not be fully represented in the training data, makes the system lack sufficient adaptability when facing these rare scenarios, thereby increasing the safety risk.

[0003] Current autonomous driving systems mainly rely on offline training methods, which require a large amount of training data to teach the system how to deal with various driving situations. However, these systems have difficulty in dealing with rare or unforeseen events, especially when these situations are not fully demonstrated in the training data. In addition, the static nature of traditional models limits their dynamic adaptability after deployment, which is crucial for achieving high adaptability and safety in the real world.

[0004] To solve these problems, the concept of "continuous learning" has emerged as a promising solution. Continuous learning enables models to adapt to new data without forgetting previously learned knowledge. The challenge lies in balancing "stability" and "plasticity".

[0005] In the prior art, the mainstream method of continuous learning still has the following significant defects when dealing with complex scenarios such as autonomous driving car-following, which have strong time series and fuzzy task boundaries. First, the means of memory regulation is too static. The mainstream regularization method (such as elastic weight consolidation) relies on static anchoring of the historical parameter space, and cannot perform fine-grained selective forgetting and updating according to the task time sequence and the contribution of the parameters to the current learning, resulting in difficulty in balancing adaptability learning and historical knowledge retention. Second, the existing methods are too single in the protection dimension of knowledge. Existing methods focus on constraining individual parameters to approach the historical optimal value, but ignore the collaborative changes of functionally related parameter groups in the neural network during continuous learning, affecting the coherent transfer of knowledge and the generalization performance of the model. Finally, the learning process of the current method is seriously disconnected from the actual task dynamics. The memory retention strength of most methods is a fixed value set in advance, and cannot be adaptively adjusted according to the difficulty of new tasks (such as the comparison of new and old task losses) or the relevance between tasks. This makes the model prone to performance fluctuations or forgetting problems when facing complex and variable continuous tasks, especially when encountering new tasks with significantly increased difficulty or low relevance. SUMMARY

[0006] In order to solve the problems in the background art, the present application provides an autonomous driving car-following control method based on brain-like continuous learning, which solves the technical problems of serious forgetting of historical knowledge, insufficient adaptation to new tasks and unstable overall control performance of existing automatic driving car-following models when continuously learning new driving scenes due to rigid memory mechanism, single knowledge protection dimension and disconnection between learning process and task dynamics.

[0007] The technical scheme adopted by the present application is:

[0008] One kind based on brain-like continuous learning's autonomous driving car-following control method:

[0009] S1, the driving data of the car-following vehicle under the incremental task is collected by the vehicle-mounted sensor, and the driving data set under the incremental task is constructed according to the driving data.

[0010] S2, the driving data set under the incremental task is input into the vehicle car-following control model trained under the previous task for continuous learning, and the vehicle car-following control model trained under the incremental task is obtained.

[0011] S3, the driving data of the car-following vehicle is collected in real time, and the driving data is input into the vehicle car-following control model trained under the incremental task for processing to obtain real-time predicted driving data.

[0012] S4, the car-following vehicle is fed back according to the real-time predicted driving data, and the car-following behavior is adjusted according to the predicted driving data.

[0013] S5, the driving data of the car-following vehicle after adjusting the car-following behavior is collected, and the collected driving data is used as the driving data of the next task.

[0014] The step S1 is specifically:

[0015] S11, the continuous time frame driving data of several car-following vehicles under the incremental task is collected by the vehicle-mounted sensor.

[0016] S12, for each car-following vehicle, the continuous time frame driving data is divided into first time period driving data and second time period driving data in a preset proportion in time sequence, the first time period driving data is used as input driving data, and the second time period driving data is used as label driving data, so as to construct a single driving data sample.

[0017] The driving data samples corresponding to the several car-following vehicles are summarized to obtain the driving data set under the incremental task.

[0018] Each row driving data of the following vehicle includes a speed of the following vehicle, an acceleration of the following vehicle, a speed of the followed vehicle, an acceleration of the followed vehicle and a distance between the two vehicles.

[0019] The vehicle following control model comprises a plurality of time series prediction modules and an integration module; each time series prediction module adopts an LSTM network; and an MSE loss function is used in the training process of each LSTM network.

[0020] The process of continuous learning of the vehicle following control model is specifically as follows:

[0021] D1, the driving data set under the incremental task is input into each time series prediction module trained in the previous task to obtain a plurality of trained time series prediction modules, and each trained time series prediction module is frozen.

[0022] D2, the driving data set under the incremental task is input into each frozen time series prediction module, each frozen time series prediction module outputs corresponding predicted driving data, all predicted driving data are input into the integration module for integration processing to obtain integrated predicted driving data, and the integration module is continuously updated according to the integrated predicted driving data, label driving data in the driving data set and a loss function, so as to obtain a trained vehicle following control model under the incremental task.

[0023] The integration module is set according to the following formula:

[0024] I = Σ i=1 o (g i ·X i )

[0025] g i = softmax(ResNet1D([X1,…,X n ]))

[0026] Wherein, I is the integrated predicted driving data obtained by the integration module; i is an index; o is the total number of time series prediction modules; g i is the output contribution weight of the i-th time series prediction module; X i represents the output of the i-th time series prediction module; X1, X i and X n represent the outputs of the first, i-th and n-th time series prediction modules respectively.

[0027] The loss function of the integration module is set according to the following formula:

[0028] J LA (θ LA )= L BLA +Ω MSUM LA (θ LA )+Ω GMMFM LA (θ LA )+Ω H (g)

[0029] L B LA =(1 / n)Σ i=1 n (I i -Y i ) 2

[0030] Ω MSUM LA (θ LA )=λ MSUM Σ t<T Σ m=1 d (α t ·ζ e,m |θ LA,m -θ e,m |)

[0031] ζ e,m =((|▽ θLA,m ·L e |) / (Σ m=1 d |▽ θLA,m L e |))·(L e / L past )

[0032] Ω GMMFM LA (θ LA )=λ GMMFM Σ t<T Σ m=1 d [γ GMMFM ζ m (θ LA,m -θ LA,m * ) 2 +Σ j=1 d λ m,j |θ LA,m -θ LA,j |]

[0033] ζ m =(|▽ θLA,m ·L past |) / (Σ m=1 d |▽θLA,m ·L past |)

[0034] Ω H (g)=-λ H Σ i=1 o (g i log(g i ))

[0035] Among them, J LA (θ LA ) represents the total loss function of the integration module; θ LA For the set of parameters in the integration module; L B LA Ω is the task loss function; MSUM LA (θ LA ) represents the memory-selective update loss function; Ω GMMFM LA (θ LA ) represents the global memory transfer and feedback loss function; Ω H (g) represents the weight entropy regularization loss function; n is the sample size of a single training batch; i is the index; I i Y represents the predicted driving data corresponding to the i-th sample; i Let λ be the label driving data corresponding to the i-th sample; MSUM The weights are selectively updated for memory purposes; t is the index, representing the t-th historical task; α t Let be the temporal preservation coefficient of the t-th historical task; T be the index of the incremental task; m be the index, m∈{1,…,d}, and d be the parameter set θ. LA The total number of trainable parameters in θ; LA,m θ represents the m-th parameter value in the integration module during the incremental task training process. e,m To determine the m-th parameter value in the empty network during the incremental task training process when the integrated module is replaced with an empty network; ζ e,m For the migration weights of incremental tasks; L e For the loss during incremental task training in an empty network, the loss function of the empty network is the task loss function; L past The optimal loss in the historical task; |▽ θLA,m ·L e | is the parameter θ in the incremental task LA,m The magnitude of the gradient; Σ m=1 d |▽ θLA,m ·L e | represents the sum of gradients for all parameters during the incremental task training process; λ GMMFM To record the hyperparameters of global memory transfer and feedback; γ GMMFMθ is a global memory transfer strength control hyperparameter LA,m * θ is the mth parameter value corresponding to the optimal loss in the historical task; j represents an index, j is an element of the set {1,..., d}; and m,j θ is the feedback transfer factor between the mth and jth parameters in the integration module LA,j θ is the jth parameter value of the incremental task training process in the integration module m θ is the transfer weight of the historical task H g represents an entropy regularization coefficient i θ is the output contribution weight corresponding to the ith time series prediction module; log is a logarithmic function; and o is the total number of time series prediction modules.

[0036] The step S3 specifically includes:

[0037] S31, collect N frames of driving data continuously from N-1 frames before the current moment to the current moment.

[0038] S32, input the continuous N frames of driving data into each time series prediction module in the vehicle following control model, and each time series prediction module outputs corresponding predicted driving data.

[0039] S33, all predicted driving data are input into the integration module for processing to obtain integrated predicted driving data, and the integrated predicted driving data are the final real-time predicted driving data.

[0040] II. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0041] III. A computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the above method.

[0042] The innovation of the present application lies in processing the driving data of the following vehicle, introducing a bionic mechanism of the brain neural system, simulating the structure and memory and forgetting mechanism of the fruit fly olfactory nervous system, designing a multi-level continuous learning architecture, and combining the memory selective update mechanism, the global memory transfer and feedback mechanism and the weight entropy regularization term mechanism proposed in the present application, effectively improving the continuous learning ability of the model, improving the adaptability of the model in the long tail event, realizing the continuous learning and online optimization of the autonomous driving system in the dynamic traffic environment, and improving the accuracy and stability of the vehicle following behavior, which has stronger practical significance than the traditional static model.

[0043] The present application has the following advantages:

[0044] This invention employs a multi-layered continuous learning architecture based on a brain-like continuous learning mechanism. The incremental learning method enables the system to learn and optimize online in real time, enhancing its ability to cope with long-tail events and complex scenarios. Simultaneously, by combining a selective memory update mechanism and a global memory transfer and feedback mechanism, it effectively balances learning new tasks with remembering old tasks, avoiding catastrophic forgetting. This invention enhances the generalization ability and stability of autonomous driving systems, significantly improves the adaptability of vehicle following control systems in dynamic traffic environments, and provides a safer and more efficient autonomous driving solution. Attached Figure Description

[0045] Figure 1 This is a flowchart of the present invention.

[0046] Figure 2 A flowchart for continuous learning.

[0047] Figure 3 This is a schematic diagram of the biomimetic architecture of the multi-level continuous learning model inspired by the fruit fly brain in Example 1.

[0048] Figure 4 This is a diagram showing the results of continuous learning of the vehicle following control model in Example 1.

[0049] Figure 5 The above are simulated speed heatmaps of the 25-vehicle convoy following scenario in Example 1 and Comparative Examples 1-3. Detailed Implementation

[0050] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0051] Example 1:

[0052] like Figure 1 As shown, the autonomous driving following control method of this embodiment includes the following steps:

[0053] S1. Collect driving data of following vehicles under incremental tasks through on-board sensors, and construct a driving dataset under incremental tasks based on the driving data.

[0054] like Figure 2 As shown, the continuous learning process is divided into historical tasks and incremental tasks. Historical tasks are the tasks that the model has learned and trained on, while incremental tasks are the tasks that the current model needs to learn.

[0055] S11, collecting continuous time frame driving data of several following vehicles under the incremental task through the vehicle-mounted sensor of the vehicle.

[0056] Each frame driving data of each following vehicle includes the speed of the following vehicle, the acceleration of the following vehicle, the speed of the followed vehicle, the acceleration of the followed vehicle, and the distance between the two vehicles.

[0057] S12, for each following vehicle, the continuous time frame driving data is divided into first time period driving data and second time period driving data in a time sequence in a preset proportion, the first time period driving data is taken as input driving data, and the second time period driving data is taken as label driving data, so as to construct a single driving data sample.

[0058] S13, the driving data samples corresponding to the several following vehicles are summarized to obtain a driving data set under the incremental task.

[0059] S2, inputting the driving data set under the incremental task into the vehicle following control model based on the brain-like continuous learning trained under the previous task for continuous learning, to obtain a vehicle following control model trained under the incremental task; the vehicle following control model comprises several time sequence prediction modules and an integration module.

[0060] In the embodiment, a multi-level continuous learning architecture is designed to imitate the structure and function of the gamma MB system in the fruit fly nervous system. It includes an integration module and multiple parallel time sequence prediction modules, and each module has independent learning ability. As shown in Figure 3 The embodiment is inspired by the five-division parallel processing structure of the gamma MB system in the fruit fly brain, and a multi-level continuous learning model bionic architecture (vehicle following control model) is designed. The architecture simulates the mechanism of cooperative work of different functional divisions in the biological nervous system, which is composed of five parallel LSTM networks (denoted as L B –L F ) and an integration module L A Each LSTM learning module learns driving data in different feature dimensions independently, and each module has an independent forgetting rate to simulate the differential regulation of DPM neurons in the fruit fly brain to the gamma division; the integration module L A ResNet (1D) is used as the specific implementation of the weighted integration network, which integrates the outputs of the five learning modules according to the optimal weight value obtained by the core training network. The MSE loss function is used in the training process of each LSTM network.

[0061] The process of continuous learning of the vehicle following control model is as follows:

[0062] D1, the driving data set under the incremental task is input into each time series prediction module trained in the last task to obtain a plurality of trained time series prediction modules, and each trained time series prediction module is frozen.

[0063] D2, the driving data set under the incremental task is input into each frozen time series prediction module, each frozen time series prediction module outputs corresponding predicted driving data, all predicted driving data are input into the integration module for integration processing to obtain integrated predicted driving data, and the integration module is continuously updated according to the integrated predicted driving data, the label driving data in the driving data set and the loss function, so as to obtain the vehicle following control model trained under the incremental task.

[0064] The target of the integration module is to realize the optimal cooperation between different learning modules through weight fusion, while maintaining the balanced learning ability of the model to the incremental task and the historical task. The integration module is set according to the following formula:

[0065] I = Σ i=1 o (g i ·X i )

[0066] g i = softmax(ResNet1D([X1,…,X n ]))

[0067] Wherein, I is the integrated predicted driving data obtained by the integration module; i is the index; o is the total number of time series prediction modules; g i is the output contribution weight of the i-th time series prediction module, which represents the output contribution weight of the i-th time series prediction module after the output of all predicted driving data is output by ResNet(1D) and normalized by softmax; softmax() represents the softmax function; ResNet1D([]) represents a one-dimensional ResNet network model; X i represents the output of the i-th time series prediction module; X1, X i and X n represent the outputs of the first, i-th and n-th time series prediction modules, respectively.

[0068] The loss function of the integration module is set according to the following formula:

[0069] J LA (θ LA )=L B LA +Ω MSUM LA (θ LA )+Ω GMMFMLA (θ LA )+Ω H (g)

[0070] L B LA =(1 / n)Σ i=1 n (I i -Y i ) 2

[0071] Ω MSUM LA (θ LA )=λ MSUM Σ t<T Σ m=1 d (α t ·ζ e,m |θ LA,m -θ e,m |)

[0072] ζ e,m =((|▽ θLA,m ·L e |) / (Σ m=1 d |▽ θLA,m ·L e |))·(L e / L past )

[0073] Ω MSUM LA (θ LA The loss function ζ is a memory-selective update function that simulates the selective retention and forgetting mechanism of memory regulated by Rac1 activity in the fruit fly brain. It dynamically allocates parameter update intensity based on parameter importance and the relative difficulty of historical task loss and incremental task loss. Parameters with larger gradients are considered more important to the current incremental task and given higher transfer weights. e,m This allows for more significant updates; simultaneously, the greater the ratio of incremental task loss to historical task loss, the stronger the overall migration. Temporal preservation coefficient α t This approach incorporates the biological principle of memory decay over time, resulting in a lower weighting of historical tasks on current parameters. Its advantages include: achieving refined parameter-level memory management by differentially forgetting old memories based on the importance of parameters in the current task and the temporal relationship of tasks, thus more effectively mitigating catastrophic forgetting. Simultaneously, it enhances the model's adaptability to incremental tasks by focusing on targeted updates to parameters important for the current task, accelerating the model's convergence speed on new tasks.

[0074] Ω GMMFMLA (θ LA )=λ GMMFM Σ t<T Σ m=1 d [γ GMMFM ζ m (θ LA,m -θ LA,m * ) 2 +Σ j=1 d λ m,j |θ LA,m -θ LA,j |]

[0075] ζ m =(|▽ θLA,m ·L past |) / (Σ m=1 d |▽ θLA,m ·L past |)

[0076] Ω GMMFM LA (θ LA The function is a global memory transfer and feedback loss function, which mimics the memory integration and consolidation process across brain regions and time scales in the fruit fly brain. This function constrains the historically optimal parameters, with the constraint strength determined by the importance ζ of the parameters in the historical task. m and global memory transfer strength coefficient γ GMMFM This collaborative decision-making process protects important historical knowledge from being overwritten. It also introduces feedback transfer constraints between parameters, using a feedback transfer factor λ. m,j This method measures and constrains the consistency of changes in functionally coupled parameter pairs (m,j) during incremental learning, simulating the cooperative adaptation characteristics of functionally related neuron groups in neural networks. Its advantages include: achieving global knowledge consolidation by explicitly anchoring historically optimal parameters, effectively combating historical performance degradation caused by learning new tasks. Simultaneously, it emphasizes cooperative transfer between parameters; the feedback transfer term constrains not only individual parameters but also the relationships between parameters, enabling the model to maintain the beneficial associations learned within or between functional modules when adapting to new tasks, enhancing the coherence of knowledge transfer and the model's generalization ability.

[0077] Ω H (g)=-λ H Σ i=1 o (g i log(g i ))

[0078] Ω H(g) is a weight entropy regularization term loss function, which aims to prevent the "information collapse" phenomenon that the weight distribution of the integration module is excessively concentrated on one or a few modules when the integration module weights the fusion of the outputs of each time series prediction module. By minimizing the negative entropy of the weight distribution g, the weight vector is encouraged to tend to be uniformly distributed. The advantages are: promoting the diversity of module utilization, ensuring that all parallel time series prediction modules can contribute to the final decision, avoiding the insufficient utilization of model capacity caused by the dominance of individual modules. And improve the robustness of the model, the dispersion of the weight can reduce the risk of overfitting of the model to a single module or a specific feature dimension, thereby improving the stability of the model when facing out-of-distribution data or noise.

[0079] wherein J LA (θ LA ) is the total loss function of the integration module; θ LA is the set of parameters in the integration module; L B LA is the task loss function; Ω MSUM LA (θ LA ) is the memory selective update loss function; Ω GMMFM LA (θ LA ) is the global memory transfer and feedback loss function; Ω H (g) is a weight entropy regularization term loss function; n is the sample size of a single training batch; i is the index; I i is the predicted driving data corresponding to the i-th sample in the training batch; Y i is the label driving data corresponding to the i-th sample in the training batch; λ MSUM is a preset memory selective update weight (a hyperparameter controlling the forgetting strength); t is an index representing the t-th historical task; α t is the time series retention coefficient of the t-th historical task, which reflects the biological law of memory decay over time; T is the index of the incremental task; m is an index, m∈{1,…,d}, d is the total number of all trainable parameters in the parameter set θ LA ; θ LA,m is the m-th parameter value of the integration module in the incremental task training process; θ e,m is the m-th parameter value of the empty network in the incremental task training process when the integration module is replaced by the empty network; ζ e,m is the transfer weight of the incremental task, representing the strength of the m-th parameter value in the incremental task that needs to be updated, which is determined based on the importance of the parameter in the incremental task; L e is the loss of the incremental task training process in the empty network, and the loss function of the empty network uses the task loss function; L past is the optimal loss in the historical task; |▽ θLA,m ·Le | is the gradient of the parameter θ in the incremental task, indicating the importance of the parameter in the incremental task, and a larger gradient indicates that the parameter contributes more to the loss function and should be updated more intensively. LA,m m=1 d θLA,m ·L e | is the sum of the gradients of all parameters in the incremental task training process, used to normalize the gradient of a single parameter. e / L past is the ratio of the loss of the incremental task to the loss of the historical task, indicating the difficulty of the incremental task. If the loss of the incremental task is larger, the transfer weight is larger, and vice versa. GMMFM is a global memory transfer and feedback hyperparameter; γ GMMFM is a hyperparameter that controls the strength of global memory transfer; θ LA,m * is the mth parameter value corresponding to the optimal loss in the historical task, used to depict the existing knowledge that should be retained; j represents the index, j ∈ {1, …, d}, and d is the total number of trainable parameters in the parameter set θ LA ; λ m,j is the feedback transfer factor between the mth and jth parameters in the integration module during the incremental task training process, used to measure the coupling strength between the mth and jth parameters in terms of function; θ LA,j is the jth parameter value in the integration module during the incremental task training process; ζ m is the transfer weight of the historical task, indicating the strength of the mth parameter value that needs to be remembered in the historical task; λ H represents the entropy regularization coefficient, used to adjust the balance of the weights of each module to prevent information collapse caused by excessively high weights of a single module, and to ensure the diversity and robustness of model learning; g i is the output contribution weight corresponding to the ith time series prediction module; log is the logarithmic function; and o is the total number of time series prediction modules.

[0080] In this embodiment, the memory forgetting mechanism is inspired by the memory forgetting mechanism in the fruit fly brain regulated by the small G protein Rac1 activity. The influence of Rac1 activity on the retention time of memory is simulated, and a memory selective update loss function and a global memory transfer and feedback loss function are added to the loss function of the model. By adjusting the active forgetting rate, the relationship between new task learning and historical task memory is balanced, thereby avoiding catastrophic forgetting and enhancing the adaptability of the model. The memory selective update loss function can select the appropriate forgetting rate of the historical memory according to the relevance and importance of the input data, wherein the forgetting rate β is reflected by the following formula:

[0081] λ MSUM =β / (1-β) ​​

[0082] Memory selective update loss function and global memory transfer and feedback loss function through λ GMMFM and λ MSUM Two hyperparameters form a complementary regulatory mechanism. Specifically, λ GMMFM is derived directly from: MSUM

[0083] λ GMMFM =1-λ MSUM

[0084] Further:

[0085] In training the benchmark model (model 0) corresponding to the vehicle following control model, the integration module adopts L B LA +Ω H (g) loss.

[0086] S3, collect driving data of the following vehicle in real time, input the driving data into the vehicle following control model trained under the incremental task to obtain predicted driving data.

[0087] S31, collect N frames of continuous driving data from N-1 frames before the current time to the current time.

[0088] S32, input the continuous N frames of driving data into each time series prediction module in the vehicle following control model, and each time series prediction module outputs corresponding predicted driving data.

[0089] S33, all predicted driving data are input into the integration module for processing to obtain integrated predicted driving data, which is the final real-time predicted driving data.

[0090] S4, feed back the current real-time predicted driving data to the following vehicle, and the following vehicle adjusts the following behavior according to the predicted driving data.

[0091] S5, collect driving data of the following vehicle after adjusting the following behavior in real time, and collect the driving data as driving data of the next task.

[0092] Further, after real-time prediction and collection for a period of time, the current following vehicle is not predicted, the next task is a new incremental task and steps S1-S2 are repeated to obtain a new vehicle following control model trained under the new incremental task, and then steps S3-S5 are executed, and the cycle is repeated, realizing continuous learning.

[0093] The embodiment also provides a computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the above method when executing the computer program.​

[0094] The embodiment also provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the steps of the method.

[0095] In the embodiment, experiments are performed on highD data collected on an open highway in a certain country and data of a following vehicle collected on a 3.2 km section of an Innovation Avenue in a certain region on January 19, 2013. The total time of collection is 269 minutes, and the frame rate of collection is 10 frames per second. Data including the speed and acceleration of the following vehicle and the speed and acceleration of the followed vehicle and the distance between the two vehicles are obtained.

[0096] In the embodiment, the first time period driving data and the second time period driving data are respectively 50 frames, k is 10 frames of driving data for a total of 6 seconds. A total of 439 million data points of more than 110,000 following vehicles are collected, of which 40 million data points are used for training of model 0 (baseline model), and 3.9 million data points are used for continuous learning.

[0097] In the embodiment, highD data collected on a highway in a certain country is used as training of model 0 (baseline model), and data of a platoon following experiment performed in a certain region is used as data of an incremental task. The experiment is divided into 3 groups, each group of experiment is completed by a following vehicle fleet consisting of 25 vehicles, and a total of 12 one-way driving tasks are completed. Each one-way driving contains three typical driving scenes of sudden acceleration, uniform speed driving and sudden deceleration. Different speeds are set in the uniform speed driving stage of each one-way driving to increase data diversity.

[0098] As Figure 4 Figure 36 shows the results of the vehicle following control model in the embodiment under 36 tasks. Different colored lines represent the iteration process of the model, and the lighter the line color, the later the iteration of the model. With the training of the incremental task, the following results of the model gradually approach the real driving data and gradually remain stable. That is, with the advancement of the continuous learning process, the following level of the model is continuously improved and remains stable, representing that the model has the ability of continuous learning and good plasticity, and can achieve smooth operation in the road environment.

[0099] Comparative Example 1

[0100] The vehicle following control model in the comparative example is trained by using the same method as that in the embodiment 1. The vehicle following control model in the comparative example is the model of the application, and only the baseline model (model 0) is trained without subsequent continuous learning of the incremental task.

[0101] The vehicle following control model in the comparative example is trained by using the same method as that in the embodiment 1. The vehicle following control model in the comparative example is the model of the application, and only the baseline model (model 0) is trained without subsequent continuous learning of the incremental task. Figure 5 (c) shown in Figure 36, and the consistency index with the real data is shown in Table 1. ByFigure 5 As shown in Table 1, the performance of the vehicle following control model in this comparative example is far lower than that of the vehicle following control model used in Example 1.

[0102] Comparative Example 2:

[0103] This comparative example uses the same method and dataset as Example 1 to conduct a 25-vehicle platoon car-following simulation experiment. However, in this comparative example, the vehicle car-following control model adopts a reinforcement learning (RL) model.

[0104] The comparative vehicle following control model's simulation results in a 25-vehicle platoon are as follows: Figure 5 As shown in (d), the consistency indicators with the real data are compared in Table 1. Through Figure 5 As shown in Table 1, the performance of the vehicle following control model in this comparative example is far lower than that of the vehicle following control model used in Example 1.

[0105] Comparative Example 3:

[0106] This comparative example uses the same method and dataset as Example 1 to conduct a 25-vehicle platoon car-following simulation experiment. However, in this comparative example, the vehicle car-following control model adopts a physical model-driven intelligent driver model (IDM).

[0107] The consistency index of the vehicle following control model in this comparative example with the real data in the 25-vehicle queuing simulation is shown in Table 1. Table 1 shows that the performance of the vehicle following control model in this comparative example is far lower than that of the vehicle following control model used in Example 1.

[0108] As shown in Table 1, the consistency index between the simulated and real car-following scenarios of each model queue reveals that the simulation results of the brain-inspired continuous learning model of this invention are close to the real data, significantly outperforming the initial model, reinforcement learning model, and IDM model. These results demonstrate that the biomimetic continuous learning model can effectively support car-following incremental task learning and is significantly superior to traditional data-driven methods and physical models in terms of plasticity.

[0109] Table 1

[0110]

[0111] Depend on Figure 5 A comparison of the speed heatmap results of the 25-vehicle platoon following scenario simulations of the various models shows that the method of this invention can enable the model to maintain the integrity of existing knowledge while dynamically adapting to new tasks through a multi-level error correction mechanism, significantly improving the stability of the vehicle following system. Its performance advantages are particularly prominent in the rear of the platoon and in long-term following scenarios.

[0112] Comparative Example 4:

[0113] The comparative example was implemented using the corresponding method of Example 1, but the loss function of the integration module was set as: J LA (θ LA )1=L B LA The results obtained are shown in Table 2.

[0114] Comparative Example 5:

[0115] The comparative example was implemented using the corresponding method of Example 1, but the loss function of the integration module was set as: J LA (θ LA )2=L B LA +Ω MSUM LA (θ LA ), the results obtained are shown in Table 2.

[0116] Comparative Example 6:

[0117] The comparative example was implemented using the corresponding method of Example 1, but the loss function of the integration module was set as: J LA (θ LA )3=L B LA +Ω GMMFM LA (θ LA ), the results obtained are shown in Table 2.

[0118] Comparative Example 7:

[0119] The comparative example was implemented using the corresponding method of Example 1, but the loss function of the integration module was set as: J LA (θ LA )4=L B LA + Ω H (g), the results obtained are shown in Table 2.

[0120] Comparative Example 8:

[0121] The comparative example was implemented using the corresponding method of Example 1, but the loss function of the integration module was set as: J LA (θ LA )5=L B LA +Ω MSUM LA (θ LA )+Ω GMMFM LA (θ LA ), the results obtained are shown in Table 2.

[0122] Comparative Example 9:

[0123] The comparative example adopts the corresponding method of Example 1, but the loss function of the integration module is set as: J LA (θ LA )6=L B LA +Ω GMMFM LA (θ LA )+Ω H (g), and the results are shown in Table 2.

[0124] Comparative Example 10:

[0125] The comparative example adopts the corresponding method of Example 1, but the loss function of the integration module is set as: J LA (θ LA )7=L B LA +Ω MSUM LA (θ LA )+Ω H (g), and the results are shown in Table 2.

[0126] Table 2

[0127]

[0128] The autonomous driving following control method based on brain-like continuous learning proposed by the present application is inspired by the neural mechanisms of biological organisms, particularly the fruit fly brain, which exhibits excellent adaptability and memory retention ability through active forgetting and memory plasticity. By simulating these biological processes, the method enables autonomous driving vehicles to continuously learn and adapt in dynamic driving environments, especially in the modeling of vehicle following behavior. The method integrates multiple time series prediction modules with different forgetting rates, allowing flexible incorporation of new driving scenarios while ensuring the stability of learned tasks.

[0129] The above examples are only preferred embodiments for fully illustrating the present application, and the protection scope of the present application is not limited thereto. Any equivalent replacement or transformation made by those skilled in the art based on the present application is within the protection scope of the present application. The protection scope of the present application is subject to the claims.

Claims

1. An autonomous driving car-following control method based on brain-like continuous learning, characterized in that, Includes the following steps: S1. Collect driving data of the following vehicles under the incremental task through vehicle-mounted sensors, and construct the driving dataset under the incremental task based on the driving data; S2. Input the driving dataset under the incremental task into the vehicle car-following control model trained in the previous task for continuous learning, and obtain the vehicle car-following control model trained under the incremental task. The vehicle following control model includes several timing prediction modules and an integration module; each timing prediction module uses an LSTM network. The specific process of continuous learning for the vehicle car-following control model is as follows: D1. The driving dataset under the incremental task is input into each time series prediction module that has been trained in the previous task for training, resulting in several trained time series prediction modules. Each trained time series prediction module is then frozen. D2. The driving dataset under the incremental task is input into each frozen time series prediction module. Each frozen time series prediction module outputs the corresponding predicted driving data. All predicted driving data are then input into the integration module for integration processing to obtain integrated predicted driving data. The integration module is continuously updated based on the integrated predicted driving data, the labeled driving data in the driving dataset, and the loss function, thereby obtaining the vehicle car-following control model trained under the incremental task. The loss function of the integration module is set according to the following formula: J LA (i LA )=L B LA +Oh MSUM LA (i LA )+Ω GMMFM LA (i LA )+Ω H (g) L B LA =(1 / n)Σ i=1 n (I i -Y i ) 2 Oh MSUM LA (i LA )=λ MSUM S t<T S m=1 d (a t ·g e,m |θ LA,m -θ e,m |) g e,m =((|▽ θLA,m ·L e |) / (S m=1 d |▽ θLA,m L e |))·(L e / L past ) Oh GMMFM LA (i LA )=λ GMMFM S t<T S m=1 d [c GMMFM g m (i LA,m -θ LA,m * ) 2 +S j=1 d l m,j |θ LA,m -θ LA,j |] g m =(|▽ θLA,m ·L past |) / (S m=1 d |▽ θLA,m ·L past |) Oh H (g)=-λ H S i=1 o (g i log(g i )) Among them, J LA (θ LA ) represents the total loss function of the integration module; θ LA For the set of parameters in the integration module; L B LA Ω is the task loss function; MSUM LA (θ LA ) represents the memory-selective update loss function; Ω GMMFM LA (θ LA ) represents the global memory transfer and feedback loss function; Ω H (g) represents the weight entropy regularization loss function; n is the sample size of a single training batch; i is the index; I i Y represents the predicted driving data corresponding to the i-th sample; i Let λ be the label driving data corresponding to the i-th sample; MSUM The weights are selectively updated for memory purposes; t is the index, representing the t-th historical task; α t Let be the temporal preservation coefficient of the t-th historical task; T be the index of the incremental task; m be the index, m∈{1,…,d}, and d be the parameter set θ. LA The total number of trainable parameters in θ; LA,m θ represents the m-th parameter value in the integration module during the incremental task training process. e,m The value of the m-th parameter in the empty network during the incremental task training process, α, when the integration module is replaced with an empty network. t ζ is the timing preservation coefficient; e,m For the migration weights of incremental tasks; L e For the loss during incremental task training in an empty network, the loss function of the empty network is the task loss function; L past The optimal loss in the historical task; |▽ θLA,m ·L e | is the parameter θ in the incremental task LA,m The magnitude of the gradient; Σ m=1 d |▽ θLA,m ·L e | represents the sum of gradients for all parameters during the incremental task training process; λ GMMFM To record the hyperparameters of global memory transfer and feedback; γ GMMFM θ is a hyperparameter used to control the strength of global memory transfer. LA,m * λ represents the m-th parameter value corresponding to the optimal loss in the historical task; j represents the index, j∈{1,…,d}; m,j θ is the feedback transfer factor between the m-th and j-th parameters in the integration module. LA,j ζ represents the j-th parameter value in the integration module during the incremental task training process. m For the migration weight of historical tasks; λ H g represents the entropy regularization coefficient; i The output of the i-th time series prediction module is assigned a weight; log is the logarithmic function; o is the total number of time series prediction modules; S3. Real-time acquisition of driving data of following vehicles, input of driving data into the vehicle following control model trained under incremental task for processing to obtain real-time predicted driving data. S4. Feedback the real-time predicted driving data to the following vehicles, and the following vehicles adjust their following behavior according to the predicted driving data. S5. Collect driving data of the following vehicle in real time after adjusting its following behavior, and use the collected driving data as the driving data for the next task.

2. The autonomous driving following control method based on brain-like continuous learning according to claim 1, characterized in that, Step S1 specifically involves: S11. Collect continuous time frame driving data of several following vehicles under incremental tasks through on-board sensors; S12. For each following vehicle, the continuous time frame driving data is divided into a first time period driving data and a second time period driving data in a preset ratio according to the time sequence. The first time period driving data is used as the input driving data and the second time period driving data is used as the label driving data, thereby constructing a single driving data sample. S13. The driving data samples corresponding to several following vehicles are summarized to obtain the driving dataset under the incremental task.

3. The autonomous driving following control method based on brain-like continuous learning according to claim 2, characterized in that: Each frame of driving data for the following vehicle includes the speed of the following vehicle, the acceleration of the following vehicle, the speed of the vehicle being followed, the acceleration of the vehicle being followed, and the distance between the two vehicles.

4. The autonomous driving following control method based on brain-like continuous learning according to claim 1, characterized in that: The integration module is configured according to the following formula: I=S i=1 o (g i ·X i ) g i =softmax(ResNet1D([X1,…,X n ])) Where I represents the integrated predicted driving data obtained by the integration module; i is the index; o is the total number of time-series prediction modules; g i The output of the i-th time series prediction module is assigned a weight; X i Represents the output of the i-th time series prediction module; X1, X i and X n These represent the outputs of the 1st, 1st, and 1st time series prediction modules, respectively.

5. The autonomous driving following control method based on brain-like continuous learning according to claim 1, characterized in that, Step S3 specifically involves: S31. Collect N consecutive frames of driving data from the N-1 frames before the current time to the current time frame; S32. Input N consecutive frames of driving data into each time-series prediction module in the vehicle following control model, and each time-series prediction module outputs the corresponding predicted driving data. S33. All predicted driving data are processed in the integration module to obtain integrated predicted driving data, which is the final real-time predicted driving data.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images