An automatic driving hierarchical decision method for mixed traffic flow

Abstract

The application discloses a kind of automatic driving hierarchical decision methods for mixed traffic flow, belong to data processing and artificial intelligence technical field, including: based on vehicle sensor obtains the image of shooting, based on YOLOv9 algorithm the feature of the image of shooting is extracted, obtains target state information;The target state information is handled, and state matrix is obtained;Graph convolution network is constructed, and the state matrix is calculated by the graph convolution network, and the topological graph structure of time series information is obtained, based on the topological graph structure of time series information obtains vehicle driving state;PID control strategy is constructed, and automatic driving is decided based on the vehicle driving state and the PID control strategy.The hierarchical decision method of the application shows good generalization ability in various application scenarios of normal road environment.

Description

Technical Field

[0001] This invention belongs to the field of data processing and artificial intelligence technology, and in particular relates to a hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow. Background Technology

[0002] With the increasing prevalence of private cars in cities, while providing great convenience for citizens' travel, it has also brought about severe traffic congestion. Daily traffic congestion leads to an exponential increase in travel time, thereby increasing driver fatigue and the risk of traffic accidents in complex traffic environments. Autonomous vehicles, capable of handling driving tasks autonomously, offer a solution to alleviate human driving stress and improve safety and accuracy, especially in monotonous traffic scenarios. The decision-making layer unit is a key component of autonomous driving algorithms, corresponding to the human decision-making process. Current research mainly falls into two categories: learning-based and rule-based methods.

[0003] Rule-based algorithms exhibit good stability, but they face limitations when processing rich perceptual information (such as pixel information in 2D images and point cloud information from 3D radar). To address this issue, reinforcement learning-based decision-making algorithms, such as DQN, DDPG, and AC, have been extensively studied due to the information processing capabilities of their network structures. However, due to the inherent characteristics of traditional reinforcement learning algorithms, these algorithms suffer from significant uncertainties in autonomous driving tasks. Specifically, they perform well in simple scenarios but face difficulties when generalizing to diverse testing scenarios. Their applications are typically limited to specific scenarios, such as one-way streets or simple intersections, while their learning performance deteriorates in complex situations. Furthermore, these algorithms heavily rely on the ideal global perception information of V2X intelligent transportation systems, assuming that the state information of surrounding vehicles is known and accurate, and that information can be shared through globally connected communication. However, in the current traffic environment, 100% connectivity of road vehicles is unattainable, and mixed traffic flow will be the norm.

[0004] Regarding the challenges of acquiring global perception information in mixed traffic flows, end-to-end algorithms can theoretically directly process information provided by existing mature perception systems (such as cameras and LiDAR) and directly output the action information required for vehicle chassis control. However, the unpredictability of end-to-end model outputs significantly impacts its practical application, a common characteristic of complex network structures. The fundamental problem lies in seamlessly integrating different stages of the entire autonomous driving information flow into a unified and difficult-to-interpret whole, turning the model into an incomprehensible black box. Uncertain, fatal logical errors could lead to serious traffic accidents; therefore, this type of algorithm remains largely confined to theoretical research. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, the present invention provides a hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow, comprising:

[0007] Based on images captured by vehicle-mounted sensors, features of the captured images are extracted using the YOLOv9 algorithm to obtain target state information;

[0008] The target state information is processed to obtain a state matrix;

[0009] A graph convolutional network is constructed, and the state matrix is ​​calculated through the graph convolutional network to obtain the topological graph structure of the time series information. The vehicle driving state is obtained based on the topological graph structure of the time series information.

[0010] A PID control strategy is constructed to make decisions on autonomous driving based on the vehicle driving state and the PID control strategy.

[0011] Preferably, the process of extracting features from the captured image based on the YOLOv9 algorithm to obtain target state information includes:

[0012] Based on the YOLOv9 algorithm, the target state information is generated from the relevant targets in the captured image and their position information in the field of view.

[0013] Preferably, the target state information includes the identified target type, the pixel position of the target in the image, and confidence information related to the identification.

[0014] Preferably, the process of processing the target state information to obtain the state matrix includes:

[0015] Based on the number of effectively identified targets in images taken from various directions, standardized coding rules are constructed, and the target state information is processed based on the standardized coding rules to obtain a state matrix.

[0016] Preferably, the vehicle driving states include normal driving state, following vehicle state, traffic intersection driving state, and emergency braking state.

[0017] Preferably, the PID control strategy includes a horizontal PID control strategy and a vertical PID control strategy;

[0018] The lateral PID control strategy is used to generate the steering wheel angle;

[0019] The longitudinal PID control strategy is used to generate throttle and brake control signals.

[0020] Preferably, the control process of the PID control strategy includes:

[0021] The control speed threshold is set based on the experimental and observational system.

[0022] If the vehicle speed exceeds the speed threshold, the parameters of the lateral PID control strategy and the longitudinal PID control strategy are recalibrated.

[0023] Compared with the prior art, the present invention has the following advantages and technical effects:

[0024] In the perceptual information preprocessing stage, this invention integrates a pre-trained target detection model into the data encoder, transforming complex image data into a state matrix beneficial to the decision network. By adding directional coordinate information to the state matrix, the algorithm's adaptability to perceptual information and its ability to understand scene details are enhanced. A state machine based on a temporal graphical convolutional network (GCN) is introduced: a time-series-based GCN model is incorporated to fuse and align the concept of time with real-time state information within the driving scenario. Compared to traditional convolutional neural networks (CNNs), GCNs excel in capturing temporal relationships; this key improvement significantly enhances the algorithm's scene understanding capabilities. Attached Figure Description

[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0026] Figure 1 This is a schematic diagram of the layered framework according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of state space encoding according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the neural network structure according to an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of time-series graph convolution according to an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of the longitudinal PID control according to an embodiment of the present invention;

[0031] Figure 6 This is a schematic diagram of the horizontal PID control according to an embodiment of the present invention;

[0032] Figure 7 This is a schematic diagram of a self-built dataset multi-level directory according to an embodiment of the present invention;

[0033] Figure 8This is a schematic diagram of the recognition accuracy of the vehicle driving state machine according to an embodiment of the present invention, wherein (a) is a schematic diagram of the recognition accuracy of different network models combined with the state machine; and (b) is a detailed comparison diagram of the two optimal groups.

[0034] Figure 9 This is a schematic diagram of the convergence speed of an embodiment of the present invention, wherein (a) is a schematic diagram of the convergence trend of training loss for different network model combinations; and (b) is a detailed comparison diagram of the two optimal groups. Detailed Implementation

[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0037] Example 1

[0038] like Figure 1 As shown, this embodiment provides a hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow, including:

[0039] 1. Input layer:

[0040] Model-based image preprocessing:

[0041] Some existing image recognition algorithms can extract all target types and their occupied pixel regions from an image, which corresponds to visual information preprocessing in autonomous driving. YOLO is a mature multi-target recognition algorithm that provides multiple pre-trained models. The latest version, YOLO v9, is currently the state-of-the-art (SOTA) for object detection. Using this model for image preprocessing, relevant targets in the image and their location information within the field of view can be directly extracted.

[0042] State matrix encoding:

[0043] Image data processed by YOLO is transformed into a series of state information different from the original pixels. This information includes the identified target type, the target's pixel position in the image, and confidence information related to the identification. Furthermore, based on the input image number, camera information corresponding to each target can be distinguished, which in turn represents the target's directional position relative to the vehicle. To facilitate the subsequent input process of the network, it is necessary to establish standardized encoding rules to generate the state matrix. The feature matrix consists of feature vectors in a unified format obtained from model preprocessing and image encoding. The classification matrix represents the representation matrix of the neural network classification result, i.e., the state machine. The state machine information represents the several driving states that the vehicle needs to be divided into in an urban scene according to a predetermined route, corresponding to different control parameters in the next layer. Driving actions are information that can be directly executed by the vehicle, such as steering wheel angle, accelerator, and brake, generated by the proportional-integral-derivative (PID) controller based on the predetermined driving route.

[0044] The encoding of the state matrix depends on the number of targets effectively identified from images captured from various directions. A high-density traffic scenario is simulated by introducing 100 autonomous NPC vehicles and 50 NPC pedestrians into a town map on the Carla simulation platform. The number of targets identified by the intelligent vehicles from four cameras was statistically analyzed over a 5-hour simulation period, resulting in 18,000 frames of data. The calculated average number of targets in the images was 10.107. Considering that the environment in front of the vehicle has a much greater impact on driving behavior than the environment behind, the five targets from the rear were allocated to the front. Each row in the dataset represents a identified target, encompassing seven key dimensions. These dimensions include the target's category assignment (ranging from 0 to 79), the precise corner coordinates of the bounding box, the confidence score of the identification, and the orientation encoding representing the target's position relative to the vehicle (ranging from 0 to 3). Based on this, the state matrix is ​​configured as follows: Figure 2 As shown.

[0045] 2. Hidden layer:

[0046] Hidden layer network structure:

[0047] The feature matrix obtained through preprocessing has significantly simplified the original image data in terms of dimension and specifications; however, it still contains a large amount of information and is not suitable for direct use in determining the driving state machine. Given the superior ability of neural networks to process large-scale information, this invention utilizes neural networks to analyze the state matrix. Since vehicle driving is continuous, historical information becomes crucial in current decisions. As part of the initial data processing, YOLO is used to extract basic information from the image data, forming the basic nodes of the topology graph. The interrelationships in the topology graph are established sequentially through a timeline as described in this paper. The image at each time stamp and its extracted feature information together constitute the nodes of the GCN. Simultaneously, the timeline, as the edge of the GCN, plays a key role in determining the content of the adjacency matrix. Therefore, this invention innovatively introduces a graph convolutional network (GCN) based on time-series information, and the overall structure of the neural network's hidden layers is as follows: Figure 3 As shown.

[0048] Time-series graph convolution:

[0049] Time series data inherently exhibit historical dependencies, meaning that the value of a variable at a given moment often depends on its previous values. This is crucial for determining the driving state machine. Graph Convolutional Networks (GCNs) can effectively capture these dependencies by incorporating information from adjacent time points within a graph structure, making them more effective at modeling temporal relationships than traditional neural networks.

[0050] GCNs can directly manipulate graphs and utilize their structural information; this concept was ported from the image domain to the graph domain. However, images typically exhibit a fixed structure, while graphs have more flexible and complex structures. The basic idea of ​​GCNs is to consider all neighboring nodes and their embedded feature information, allowing convolutional computations on the topological graph. If the information at each time step is considered as a node, the time-series information actually forms a unique topological graph structure, such as... Figure 4 As shown.

[0051] 3. Output layer:

[0052] Autonomous vehicles typically follow a predetermined route throughout the driving process to ensure they reach their destination. However, in real-world driving scenarios, vehicles are occasionally disturbed due to interactions with the external environment. These disturbances can be characterized through the classification of the driving state machine. Drawing inspiration from human drivers' behavior on urban roads, vehicle driving states can be classified as follows, and this classification also determines the dimension of the model's output layer.

[0053] Normal driving conditions: The vehicle is driven on a regular road section without traffic signal restrictions and maintains a safe distance from surrounding vehicles.

[0054] Following vehicle status: In scenarios with heavy traffic, vehicles need to follow the vehicle in front and maintain a fixed safe distance.

[0055] Traffic intersection driving status: Interactions between vehicles traveling in the opposite direction and perpendicular to the target vehicle within the target vehicle's future driving area, including all types of intersections such as crossroads, T-junctions, and roundabouts, and where traffic lights are present.

[0056] Emergency braking: The vehicle applies emergency braking just before a collision or a traffic violation (such as running a red light).

[0057] 4. Execution layer:

[0058] PID-based control strategy

[0059] The PID control process can be divided into two parts: lateral control and longitudinal control, which generate steering wheel angle and throttle / brake control signals, respectively, such as... Figure 5 and Figure 6 As shown.

[0060] During the invention process, adjusting the parameters of the PID controller was a major task. Adjusting PID parameters is an empirical process, typically involving experimentation and observation of the system's response. After the first set of PID parameters stabilized and controlled the vehicle acceleration above 50 km / h, the control effect significantly decreased. Therefore, when the speed exceeded 50 km / h, the PID parameters needed to be recalibrated. Adjusting the PID parameters ensured that the system could operate stably within a speed range of 50 km / h to 100 km / h, with the upper speed limit set referencing the maximum permissible speed on Beijing's expressways. Given that vehicle control stability is directly related to speed, this invention sets 50 km / h as the critical point and establishes two different sets of PID parameter configurations, as shown in Table 1.

[0061] Table 1

[0062]

[0063] 5. Self-build training dataset:

[0064] The learning-based algorithm model proposed in this invention requires a large amount of data for training, which necessitates data collection through experiments. Carla is a mature urban traffic scene simulator that can generate highly realistic city maps within its environment. Furthermore, it allows the definition of vehicle and pedestrian NPCs for interactive purposes. Driving state information (label values) of vehicles can be directly obtained within the simulation environment. By allowing autonomous vehicles to randomly drive on various routes within the city, a large amount of image, state matrix, and driving state machine (label value) information can be collected, creating a comprehensive dataset. Driving states are directly derived from parameter queries in the Carla simulator and are divided into four different categories, which form the basis of the aforementioned state machine classification. To ensure comprehensive database coverage, time and lighting conditions in the simulation environment are randomly varied during data collection. The dataset directory is set as follows: Figure 7 As shown, the final dataset length is 396,058.

[0065] Sensor arrangement: The arrangement of the four cameras is as follows Figure 2 As shown on the left, this ensures a field of view without blind spots.

[0066] like Figure 8-9 As shown, the beneficial effects of the present invention are as follows:

[0067] The stability and accuracy of the decision-making method are improved: Compared with traditional end-to-end autonomous driving algorithms, the hierarchical framework of this invention improves the stability of the system by processing tasks in layers. The pre-trained YOLO model and time-series GCN network significantly improve the recognition accuracy of the vehicle driving state machine, reaching nearly 90%, far exceeding the basic FCN and mainstream image processing models such as VGG16, ResNet50 and EfficientNetB7.

[0068] Efficient training process: The hierarchical framework integrates the YOLO model and temporal graph convolutional representation, making the algorithm training process more targeted, accelerating the convergence speed, significantly improving the model's pre-training efficiency and optimizing the convergence effect.

[0069] Generalization capability for various driving scenarios: This algorithm can handle autonomous driving tasks in various types of scene maps and has a superior generalization capability compared to other learning-based training models.

[0070] Carla is an open-source simulation simulator developed under the guidance of the Computer Vision Center at the Autonomous University of Barcelona, ​​Spain. It aims to help developers design, train, and evaluate autonomous driving systems and has the following features:

[0071] Highly customizable urban environment: Carla provides an open and customizable urban environment, including elements such as roads, vehicles, pedestrians, and traffic lights, which can be adjusted and modified as needed;

[0072] Realistic simulation environment: Carla is committed to providing a highly realistic simulation environment that allows developers to test various autonomous driving algorithms and systems in a virtual world and evaluate their performance in the real world.

[0073] Open APIs and tools: Carla provides a rich set of APIs and tools that enable developers to easily interact with the simulation environment and customize and extend the platform to meet their specific needs;

[0074] Multiple sensor simulation: Carla supports the simulation of various sensors, including cameras, LiDAR, millimeter-wave radar, etc., enabling developers to test the performance of different sensor combinations;

[0075] As shown in Table 2, Carla provides more than 10 maps for direct use. This invention selects four representative maps to verify the generalization ability of the model algorithm. The map styles and data are as follows (the control group uses a rule-based algorithm combined with readily available global vehicle status information, while this invention directly uses the vehicle camera data as input and directly outputs the vehicle's action information):

[0076] Table 2

[0077]

[0078] Enhanced adaptability to the environment: The algorithm can still perform autonomous driving tasks in mixed traffic flow environments with incomplete global information, providing a safe and feasible solution.

[0079] Stronger anti-interference capability: The input information required by the algorithm is obtained through the vehicle camera and IMU, and the decision-making process is carried out entirely on the in-vehicle computing platform (without cloud assistance). The information is transmitted through cables and is not affected by the external network status, which enhances the robustness against signal interference.

[0080] Practical application potential: By using the best state-of-the-art image preprocessing model and a more powerful training hardware platform in real time, prediction accuracy and overall performance can be further improved.

[0081] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow, characterized in that, Includes the following steps: Based on images captured by vehicle-mounted sensors, features of the captured images are extracted using the YOLOv9 algorithm to obtain target state information; The target state information is processed to obtain a state matrix; A graph convolutional network is constructed, and the state matrix is ​​calculated through the graph convolutional network to obtain the topological graph structure of the time series information. The vehicle driving state is obtained based on the topological graph structure of the time series information. A PID control strategy is constructed, and autonomous driving decisions are made based on the vehicle driving state and the PID control strategy. The process of extracting features from the captured image based on the YOLOv9 algorithm to obtain target state information includes: Based on the YOLOv9 algorithm, the target state information is generated from the relevant targets in the captured image and their position information in the field of view. The target state information includes the identified target type, the pixel position of the target in the image, and confidence information related to the identification. The process of processing the target state information to obtain the state matrix includes: Based on the number of effectively identified targets in images taken from various directions, standardized coding rules are constructed, and the target state information is processed based on the standardized coding rules to obtain a state matrix.

2. The hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow according to claim 1, characterized in that, The vehicle driving states include normal driving state, following vehicle state, traffic intersection driving state, and emergency braking state.

3. The hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow according to claim 1, characterized in that, The PID control strategy includes a horizontal PID control strategy and a vertical PID control strategy; The lateral PID control strategy is used to generate the steering wheel angle; The longitudinal PID control strategy is used to generate throttle and brake control signals.

4. The hierarchical decision-making method for autonomous driving oriented towards mixed traffic flow according to claim 3, characterized in that, The control process of the PID control strategy includes: The control speed threshold is set based on the experimental and observational system. If the vehicle speed exceeds the speed threshold, the parameters of the lateral PID control strategy and the longitudinal PID control strategy are recalibrated.

Images