An adaptive perception based cross-scenario autonomous driving decision method

Abstract

This invention discloses a cross-scenario autonomous driving decision-making method based on adaptive perception. This invention effectively transforms complex environmental information into basic scenarios and makes decisions based on these scenarios. Compared to traditional autonomous driving technologies, this invention shows significant advantages in complex traffic environments. When facing unfamiliar scenarios, this invention can utilize existing training experience and transformed perception images, reducing reliance on large amounts of data and long training times. By simulating the human learning and driving process, this invention enables the autonomous driving system to make accurate decisions using existing training knowledge in unfamiliar scenarios, thus avoiding dependence on large amounts of training data and long training times. Through this adaptive perception mechanism, it can also process and transform bird's-eye view images of varying complexity, providing more easily understood input to the decision-making module, resulting in a more efficient and adaptable autonomous driving solution.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a cross-scenario autonomous driving decision-making method based on adaptive perception. Background Technology

[0002] With the rapid development of autonomous driving technology, most autonomous driving technologies rely on modular perception-decision frameworks as the eyes and brain of autonomous vehicles. This design approach aims to help autonomous vehicles understand their dynamic surroundings and make corresponding steering / speed change decisions. Both the perception and decision-making modules can be developed based on classical rule-based methods or deep learning methods, attempting to replace the human driver's observation and decision-making to prevent dangerous driving behaviors (e.g., sudden braking and lane changes). While they have achieved safe, efficient, and comfortable autonomous driving in specific test scenarios, they lack the ability to understand new scenarios with significant changes or completely unfamiliar environments, leading to decision-making system failure. This is one reason why we haven't seen this type of technology deployed in autonomous vehicles yet. Therefore, designing an autonomous driving perception and decision-making framework that can be used across various scenarios is a very challenging problem.

[0003] With the rapid development of computer vision technology, the perception module of autonomous vehicles can perceive and interpret the surrounding environment using a combination of hardware and algorithms. For example, it can perceive relevant information about the surrounding environment through cameras and LiDAR, and then use algorithms to fuse multimodal perception results into a bird's-eye view (BEV) image. Convolutional neural networks are then used to extract relevant features and obtain its low-dimensional representation. Although similar perception methods can preserve the road structure and the state of surrounding vehicles well, understanding the complex and ever-changing scenes within the field of vision of autonomous vehicles remains a challenging task. Furthermore, the decision-making stage typically utilizes information collected from the perception stage to make corresponding decisions through rule-matching algorithms or reinforcement learning-based methods. However, rule-based algorithms, such as adaptive cruise control, heavily rely on rule-matching algorithms and struggle to make reasonable decisions in complex traffic systems. While reinforcement learning-based methods can make appropriate decisions in fixed scenarios, their training experience cannot cover all possible traffic conditions, leading to incorrect decisions when encountering unfamiliar scenarios.

[0004] In general, existing methods are neither adaptable to complex scenarios nor can they be extended to unseen scenarios, which hinders the development of autonomous driving. Developing perception and decision-making frameworks that can adapt to different scenarios has always been an unresolved issue in the field of autonomous driving. One simple solution is to collect as many test scenarios as possible to train conventional perception and decision-making modules. However, this approach not only requires exploring a large number of extreme cases, but also consumes a long training time. Therefore, given limited resources and training scenarios, prioritizing the development of decision-making capabilities of autonomous vehicles in basic scenarios with simple road layouts, and then extending to more complex scenarios, is a more reasonable solution. However, this approach faces many difficulties, such as: (1) It is difficult to understand complex and ever-changing traffic scenarios through perception modules, and at the same time map these changing scenarios to a basic and interpretable space. (2) Reinforcement learning-based decision-making methods can adapt well to dynamic basic scenarios, but it is difficult to ensure that the decision results of basic scenarios are still effective in complex scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-scenario autonomous driving decision-making method based on adaptive perception.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0007] This invention includes the following steps:

[0008] S1: The adaptive perception module acquires bird's-eye view images captured in complex scenes and encodes them into scene geometric topology features and scene location distribution features, respectively;

[0009] S2: The geometric features in the complex scene are replaced with the geometric features of the basic scene by an adaptive perception module based on variational autoencoder, and then combined and decoded to generate a new BEV image.

[0010] S3: The results generated by the adaptive perception module are used as the input to the behavior decision module. The behavior decision module consists of a deep reinforcement learning model and a hybrid reward function. The behavior decision module performs excellent lane changing and speed change decisions based on the information provided by the perception module.

[0011] The beneficial effects of this invention are:

[0012] This invention presents a cross-scenario autonomous driving decision-making method based on adaptive perception. Compared to existing technologies, this invention effectively transforms complex environmental information into basic scenarios and makes decisions based on these scenarios. This unique method mimics the learning process of human drivers, starting with simple environments and using analogy strategies to handle more complex driving challenges. In the perception phase, bird's-eye view (BEV) images are used to capture and encode information from complex scenarios, further transforming it into geometric and distributional features. These features, after processing and decoding, are converted into new BEV images for the basic scenarios, enabling the autonomous driving system to quickly and accurately understand its dynamic environment. In the decision-making phase, combining a deep reinforcement learning model and a hybrid reward mechanism, the system can make lane-changing and speed adjustment strategies aimed at optimizing driving safety and efficiency in basic scenarios. This strategy ensures the applicability and accuracy of decisions in different scenarios, especially complex ones. Compared to traditional autonomous driving technologies, this invention demonstrates significant advantages in complex traffic environments. Traditional rule-based methods struggle to make correct decisions in unknown or abnormal traffic conditions, while deep learning and reinforcement learning, although effective in decision training scenarios, may still produce misjudgments in unknown traffic scenarios. However, this invention can leverage existing training experience and transformed perceptual images when facing unfamiliar scenarios, reducing reliance on large amounts of data and long training times. By simulating the human learning and driving process, this invention enables the autonomous driving system to make accurate decisions using existing training knowledge in unfamiliar scenarios, thus avoiding dependence on large amounts of training data and long training times. Through this adaptive perception mechanism, this invention can also process and transform bird's-eye view images of varying complexity, providing more easily understood input to the decision-making module. Therefore, compared to other methods, this invention provides a more efficient and adaptable autonomous driving solution. Attached Figure Description

[0013] Figure 1 This is a diagram illustrating the overall architecture of the adaptive perception-based behavioral decision-making framework of this invention.

[0014] Figure 2 This is a network structure diagram of the adaptive sensing module EVA-AIN of the present invention;

[0015] Figure 3 This is a network structure diagram of the behavioral decision-making module of the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0017] like Figure 1-3As shown: First, the adaptive perception module acquires a bird's-eye view of the complex scene and encodes it into scene geometric topology features and scene location distribution features. Then, the adaptive perception module based on a variational autoencoder replaces the geometric features in the complex scene with the geometric features of the basic scene, and combines and decodes them to generate a new BEV image. The generated BEV image simultaneously possesses the perceived content of key driving information in the complex scene and an easily understandable basic scene style. This maps the complex and ever-changing scene to a basic and interpretable space, improving the ability of autonomous vehicles to make safe decisions in unknown and unfamiliar environments. The results generated by the adaptive perception module are used as input to the behavior decision module. The behavior decision module consists of a deep reinforcement learning model and a hybrid reward function. Its purpose is to execute excellent lane-changing and acceleration decisions based on the information provided by the perception module.

[0018] During the training phase of the adaptive perception module EVA-AIN, two BEV images extracted from similar scenes are encoded as scene geometry and scene distribution features, respectively. Then, by swapping the scene geometry features between the two images and decoding them into new BEV images, a loss function is calculated between the BEV images before and after the transformation. This improves EVA-AIN's ability to disentangle perceptual information; for example, the scene geometry features encoded by EVA-AIN do not contain information related to scene distribution. After the adaptive perception module has good disentanglement capabilities for features, EVA-AIN can encode two independent spaces representing the geometric topology and road location distribution features in complex scenes. Then, the geometric topology features are replaced with the corresponding features of the basic scene and decoded, thereby mapping complex and varied scenes to a basic and easily understood space. The encoder consists of a pre-trained ResNet18 and two fully connected layers, denoted as Enc(x) = (f... z (x),f g (x)), where the characteristics of the scene road location distribution are determined by f z (x)=(μ,σ)=Z x The distribution form represents the scene's geometric topological features, which are represented by f. g (x)=G x The decoder is represented in vector form. It consists of four convolutional modules, each containing a deconvolutional layer, an AdaIN adaptive instance normalization layer, and a LeakyReLU activation layer. The convolutional module can be represented as:

[0019] DecBlock(x)=LeakyReLU(AdaIN[G x |DeconvBlocks(Z x )])

[0020] The encoder output is decoded from 128 channels to a 1-channel BEV image using four convolutional blocks, where the geometric features G... x The input is fed into the deconvolutional layer DeconvBlocks, and the location distribution features Z x The output of the previous deconvolutional layer is combined with the output of the AdaIN layer, and finally the LeakyReLU layer is used as the output of the current convolutional block.

[0021]

[0022] The L1 loss function consists of two parts: reconstruction error and KL divergence. The reconstruction error measures the difference between the original image and the generated image, where the generation result is obtained by encoding the input data x into the feature space z, g using an encoder, denoted as q. φ (z|x,g) is decoded into the generated result represented as pθ(x|z,g). On the other hand, KL divergence is used to measure the difference between the distributed feature space z obtained by the encoder and the previous spatial distribution p(z). Minimizing KL divergence ensures that the learned feature space distribution is as close as possible to the prior distribution, avoiding overfitting and improving generalization ability. The EVA-AIN module designed accordingly can effectively adapt to changes in road geometry and position distribution, ensuring consistency between the transformed BEV image and the real-world road distribution. Therefore, the positions of the road and the autonomous vehicle are subsequently used as a reference for mapping the positions of surrounding vehicles. By calculating their relative distances to the lane centerline and the autonomous vehicle, the positions of surrounding vehicles can be accurately transferred to the generated BEV image.

[0023] By utilizing an adaptive perception module, BEV images of real-world scenes can be transformed into styles resembling those of a base scene. Subsequently, the reinforcement learning-based action decision module only needs to be trained on this base scene to adapt to real-world scenarios. Therefore, the goal is to solve the single-scene MDP problem using a deep reinforcement learning model with mixed rewards, focusing on learning the optimal decision under the base scene. Specifically, the PPO algorithm with an Actor-Critic structure is used. The Actor network is responsible for selecting the best action given the current environment state, while the Critic network represents the state-value function, which, combined with the reward from the environment, is used to calculate the PPO advantage function to guide policy updates. The feature extractor of the action decision module consists of three convolutional layers that convert the input image at time t into a 64-channel feature map M. tThe feature map is flattened into a one-dimensional vector and then used as input to the Actor-Critic network. The Actor network consists of three linear layers, branching into two outputs, each generating three class probability distributions. These probability distributions are sampled to determine the optimal action for the current state. The two branches output lane change options, respectively. and speed change The decision result is given by L, which represents one of three lane change behaviors: lane left (ll), lane right (lr), and lane holding (lk) (i.e., L∈{ll, lr, lk}), and V, which represents one of three speed change behaviors: acceleration (su), deceleration (sd), and maintaining speed (sm) (i.e., V∈{su, sd, sm}). Due to the adoption of a macroscopic discrete action space, autonomous vehicles can execute decision-making actions across scenarios using a traditional controller. The Actor network computation process is as follows:

[0024]

[0025] Where φ1, φ2, and φ3 are linear layers, and b1, b2, and b3 are their biases.

[0026] The Critic network and Actor network use the same input to compute the state-value function. Their task is to predict the attainable reward in a given state. The computation process is as follows:

[0027]

[0028] Where φ4, φ5, and φ6 are linear layers, and b4, b5, and b6 are their biases.

[0029] The following section details the reward function of the behavior decision module, which evaluates the quality of autonomous vehicle behavior. It consists of three main parts: safe driving capability, driving efficiency, and interference with the surrounding environment. For the measurement of safe driving capability, the TTC (Total Traffic Contribution) index for autonomous vehicles is used for evaluation, and its calculation formula is as follows:

[0030]

[0031] In this reward function, a collision refers to a vehicle crashing into or violating road boundaries. Typically, autonomous vehicles use TTC (Traffic Tractor)... t A value less than 4 triggers a collision, indicating an exponential decline in safety.

[0032] The reward value for driving efficiency is directly measured by the lane-directional speed of the autonomous vehicle, defined as follows:

[0033]

[0034] Among them, At .v represents the current speed of the vehicle along the lane, v max and v min This indicates the maximum and minimum speed limits for the current lane.

[0035] The assessment of interference with the surrounding environment is rewarded by using the interference experienced by vehicles following the autonomous vehicle:

[0036]

[0037] Where v impact It is used to determine the threshold of impact of autonomous vehicles on following vehicles. Additionally, 2a... ′ *Δt represents the change in velocity between two consecutive time steps, used to determine whether vehicles behind the autonomous vehicle are affected.

[0038] like Figure 3 As shown, the transformation capability of the adaptive perception module EVA-AIN was evaluated, which involves mapping BEV images captured by the vehicle within a 50x50 meter range from the real scene to a basic and interpretable space. During the scene transformation process, this invention utilizes observations collected by the vehicle in a simulator to train the adaptive perception module in highway and ring road scenarios. The pre-trained adaptive perception module supports flexible mapping of the autonomous vehicle in the trained scenario, thereby helping the autonomous vehicle, which learns decisions in the basic scenario, to effectively understand other changed scenarios. For example, when the reinforcement learning-based decision-making module of the autonomous vehicle is trained in scenario A (highway), it can use the adaptive perception module to transform the driving information in the encountered scenario B (ring road) into an interpretable style of scenario A. The generated results clearly show that the lane position of the autonomous vehicle and the relative positions of surrounding vehicles remain unchanged before and after the scene transformation, ensuring that decisions based on the transformed scenario are still effective. Simultaneously, the generated results are filtered: when the pixel value range is 0-1, pixels less than 0.3 are corrected to 0, and pixels greater than 0.3 are corrected to 1 to ensure the quality of the generated results.

[0039]

[0040] This invention compares an autonomous driving training scheme without an adaptive perception module in two scenarios—a highway and a ring road—using a traffic simulator. Both methods' decision-making modules are trained using reinforcement learning in the highway scenario. The scheme with the adaptive perception module EVAIN performs well in metrics such as collision rate (NCC), lane departure rate (NCL), average arrival time (AvgT), minimum collision time (M-TTC), average speed (AvgVel), and average speed change of following vehicles (DecC). Furthermore, the decision-making model trained in the highway scenario, combined with the adaptive perception module, adapts well to changes in the scenario and is also applicable to the ring road. In contrast, the reinforcement learning decision-making model without the adaptive perception module fails to make correct decisions on the ring road.

[0041] This invention offers significant advantages. Traditional autonomous driving methods, such as rule-based matching algorithms, often struggle to make decisions in complex traffic systems. For example, adaptive cruise control and rule-based lane-changing models heavily rely on pre-set rules, and their decision-making is often limited when faced with unfamiliar or unusual traffic situations. While deep learning and reinforcement learning methods perform well in certain scenarios, their training experience cannot cover all possible traffic conditions, often leading to incorrect decisions in unknown environments. This invention eliminates the reliance on massive amounts of data and lengthy training processes, simulating human learning and driving methods. This allows autonomous driving systems to make decisions using learned basic knowledge and skills when facing unfamiliar scenarios. Furthermore, the adaptive perception module effectively processes and simplifies complex bird's-eye view images, transforming them into basic scenarios with simple road networks, thereby further optimizing the efficiency and accuracy of decision-making. In summary, this invention provides an economical, efficient, and highly adaptable autonomous driving decision-making framework.

[0042] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A cross-scenario autonomous driving decision-making method based on adaptive perception, characterized in that, Includes the following steps: S1: The adaptive perception module acquires bird's-eye view images captured in complex scenes and encodes them into scene geometric topology features and scene location distribution features, respectively; S2: The geometric features in the complex scene are replaced with the geometric features of the basic scene by an adaptive perception module based on variational autoencoder, and then combined and decoded to generate a new BEV image. S3: The results generated by the adaptive perception module are used as the input to the behavior decision module. The behavior decision module consists of a deep reinforcement learning model and a hybrid reward function. The behavior decision module performs excellent lane changing and speed change decisions based on the information provided by the perception module. The encoder of the adaptive perception module based on variational autoencoder consists of a pre-trained ResNet18 and two fully connected layers, denoted as follows: The characteristics of the distribution of road locations in the scene are determined by The distribution form represents the scene's geometric topological features, which are represented by... The vector form of ; In step S3, the feature extractor of the behavior decision module consists of three convolutional layers, which convert the input image at time t into a 64-channel feature map. The feature map is flattened into a one-dimensional vector and then used as the input to the Actor network. The Actor network consists of three linear layers, branching into two outputs, which respectively output the lane change. and speed change The decision result is given by , where is one of three lane change behaviors: lane left ll, lane right l t, and lane hold l t, and is one of three speed change behaviors: speed t d, speed t d, and speed hold l t.

2. The cross-scenario autonomous driving decision-making method based on adaptive perception according to claim 1, characterized in that: The decoder in step S2 consists of four convolutional modules, each containing a deconvolutional layer, an AdaIN adaptive instance normalization layer, and a LeakyReLU activation layer; the convolutional module can be represented as: The encoder output is decoded from 128 channels to a 1-channel BEV image using four convolutional blocks, where geometric features are incorporated. Input to deconvolution layer Location distribution characteristics The output of the previous deconvolutional layer is combined with the output of the AdaIN layer and then fed into the LeakyReLU layer as the output of the current convolutional block. loss function It consists of two parts: reconstruction error and KL divergence. The reconstruction error measures the difference between the original image and the generated image, where the generation result is obtained by using an encoder to process the input data. Encoding to feature space What is obtained is represented as And decode it into the generated result as KL divergence is used to measure the difference between the distributed feature space z obtained by the encoder and the previous spatial distribution. The differences between them.

3. The cross-scenario autonomous driving decision-making method based on adaptive perception according to claim 1, characterized in that: The Actor network computation process is as follows: in It is a linear layer. It is their bias; The Critic network and Actor network use the same input to compute the state-value function. Their task is to predict the attainable reward in a given state. The computation process is as follows: in It is a linear layer. It is their bias.

4. The cross-scenario autonomous driving decision-making method based on adaptive perception according to claim 3, characterized in that: The rewards are used to evaluate the performance of autonomous vehicles, and are mainly divided into: safe driving ability, driving efficiency, and interference with the surrounding environment. The safe driving ability is measured using the TTC index of autonomous vehicles, and its calculation formula is as follows: In this reward function, a collision refers to a vehicle colliding with or violating road boundaries; when an autonomous vehicle... A value less than 4 triggers a collision, indicating an exponential decline in safety. The reward value for driving efficiency is directly measured by the lane-direction speed of the autonomous vehicle, and is defined as follows: in, This indicates the current speed of the vehicle along the lane. and This indicates the maximum and minimum speed limits for the current lane. The evaluation reward for interference from the surrounding environment is based on the interference experienced by vehicles following the autonomous vehicle: in It is used to determine the threshold of the impact of autonomous vehicles on vehicles behind them; It represents the change in speed between two consecutive time steps and is used to determine whether vehicles behind the autonomous vehicle are affected.

Images