The automatic driving system is a comprehensive system that integrates functions such as environmental perception, decision control and action execution, and is a system that fully considers the coordinated planning of vehicles and the traffic environment, and is also an important part of the future intelligent transportation system. This paper focuses on the analysis of the relevant technologies of automatic driving decision control and explores the future development direction.
(Image from the Internet)
Introduction to Autonomous Driving Systems
In the usual sense, the automatic driving system can be divided into a perception layer, a decision-making layer, and an execution layer.
Perception layer
The perceptual layer is defined as the collection and processing of environmental information and vehicle information, involving road boundary detection, vehicle detection, pedestrian detection and other technologies, which can be considered as an advanced sensor technology, and the sensors used include lidar, camera, millimeter wave radar, ultrasonic radar, speed and acceleration sensors. Due to the limitations of perception of a single sensor, which cannot meet the precise perception of various working conditions, self-driving cars need to use multi-sensor fusion technology to achieve smooth operation in various environments, which is also the key technology of the perception layer.
Decision-making layer
The decision-making level can be understood as based on the perception of information to make decision-making judgments, determine the appropriate working model, formulate corresponding control strategies, and replace human drivers to make driving decisions. This part functions like giving a corresponding task to a self-driving car. For example, in the system of lane keeping, lane departure warning, distance keeping, obstacle warning, etc., it is necessary to predict the state of the vehicle and other vehicles, lanes, pedestrians, etc. in the future for a period of time. Advanced decision theories include fuzzy inference, reinforcement learning, neural networks, and Bayesian network techniques. Due to the variety of road conditions and scenarios faced by human driving, and the different people's driving strategies for different situations are also different, the optimization of human-like driving decision algorithms requires very complete and efficient artificial intelligence models and a large amount of effective data. This data needs to cover as many rare road conditions as possible, which is the biggest bottleneck in the development of driving decisions.
Execution layer
The execution layer refers to the system controlling the vehicle according to the decision result after making the decision. Each control system of the vehicle needs to be able to connect with the decision system through the bus, and can accurately control the degree of acceleration, braking degree, steering amplitude, lighting control and other driving actions according to the bus instructions issued by the decision system to achieve autonomous driving of the vehicle.
Figure 1 Introduction to the autonomous driving system
Introduction to decision control systems
The decision control software system of the traditional automatic driving system includes functional modules such as environmental prediction, behavior decision-making, action planning, and path planning.
Environment Prediction Module
As one of the direct data upstreams of the decision planning control module, the main role of the environmental prediction module is to predict the behavior of the objects recognized by the perception layer, and to convert the predicted results into trajectories of the time and space dimensions and pass them to the subsequent modules. The object information typically output by the perception layer includes physical properties such as position, velocity, and direction.
Using the physical properties of these outputs, it is possible to make "instantaneous predictions" of objects. The environmental prediction module is not limited to making predictions about objects in combination with physical laws, but can combine objects and surrounding environments and accumulated historical data information to make more "macro" behavioral predictions about perceived objects. For example, in Figure 2, by identifying the historical movement of pedestrians on the sidewalk, it is predicted that the traveler may cross the intersection on the sidewalk, and the historical trajectory of the vehicle can be judged to turn right at the intersection.
Figure 2 Schematic diagram of environmental prediction
Behavioral Decision Module
The behavioral decision module plays the role of "co-pilot" in the entire automatic driving decision planning and control software system. This level brings together all the important vehicle perimeter information, including not only the real-time location, speed, and direction of the autonomous vehicle itself, but also all the relevant obstacle information within a certain distance around the vehicle and the predicted trajectory. The problem that behavioral decision-makers need to solve is to decide on the driving strategy of self-driving cars based on knowing this information.
Because of the need to consider many different types of information, behavioral decision-making problems are often difficult to solve with a single mathematical model, but to use some advanced concepts of software engineering to design a rule engine system. For example, in the DARPA Challenge, Stanford's unmanned vehicle system utilized a series of cost designs and finite state machines to design the trajectory and handling instructions of the unmanned vehicle. At present, markov's model of decision-making process has also begun to be more and more applied to the implementation of decision algorithms at the behavioral level of autonomous driving systems. In short, the behavioral decision-making level needs to combine the results of the environmental prediction module and output macro decision instructions for subsequent planning modules to be more specifically executed.
Motion planning module
The autonomous vehicle planning module consists of two parts: motion planning and path planning. The action planning module is mainly for short-term or even instantaneous actions, such as turning, obstacle avoidance, overtaking and other actions; while the path planning module is the planning of the vehicle driving path for a long time, such as route design or selection between the departure point and the destination.
The design idea of the automatic driving system is to establish several driving states and trigger the switching of driving states through different conditions. This design idea has the problem of poor smoothness of the switching process. In the actual system design process, the real and non-real targets in the road are mainly described as virtual particles to enhance the smoothness of vehicle driving. Among them, the real target mainly refers to the vehicle, pedestrian and other factors; the non-real target includes speed limit, red light, parking point, road curvature, weather conditions, etc. The advantage of the virtual particle model method is that the algorithm model is unified, which effectively avoids the problem of vehicle acceleration and deceleration speed jump caused by target or control mode switching in the traditional control algorithm.
Path planning module
The path planning module of autonomous vehicles refers to the effective path that can safely reach the target point without collision and safely according to the performance indicators after giving the starting point and target point of the autonomous vehicle on the basis of a certain environmental model. Route planning consists of two main steps: establishing an environment map containing obstacle areas and free areas, and selecting an appropriate route search algorithm in the environment map to quickly and real-time search for drivable routes. The result of the route planning acts as a navigation for the vehicle, guiding the vehicle from its current location to the target location. The environment map representation method is mainly divided into metric map representation, topological map representation, etc.
Trends
The development of artificial intelligence machine learning, deep neural networks and networked communication technologies has further enriched the technical path of the development of autonomous vehicles, and also promoted the development of autonomous driving technology from a single prototype demonstration to a typical traffic scenario with certain landing application capabilities and independent positioning.
artificial intelligence
Artificial intelligence is a new technical science for researching and developing theories, methods, technologies and application systems used to simulate, extend and extend human intelligence. It is intended to explore the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. One of its important application areas is automatic driving, the main goal is to make self-driving cars have a certain degree of independent learning ability, and can form a memory cognition of simple traffic environment, the main application of artificial intelligence technology in the field of autonomous vehicles at this stage is reflected in the following aspects.
1. Realize the recognition and cognition of environmental objects
Using multi-eye vision, lidar, millimeter-wave radar and other sensor devices and identification algorithms, it is possible to accurately identify multi-surface objects in the actual road environment. At the same time, after the deep learning technology is integrated, iterative classification can be formed for the three-dimensional spatial dimensions and feature information of each object, so that the autonomous vehicle has the ability to recognize and recognize a variety of environmental objects.
2. Realize the detection of the drivable area
The use of advanced sensor-based map acquisition technology can extract detailed labeling (signs, markings, signal lights, etc.) and high-precision location (longitude, latitude, altitude, etc.) and other information of the road, so as to realize the extraction of road plane features by autonomous vehicles, and at the same time, based on deep learning, cognitive recognition of road drivable and non-drivable areas can be realized.
3. Realize the planning and decision-making of driving paths
Decision planning processing is another important application scenario of artificial intelligence technology in autonomous driving. At this stage, the mainstream artificial intelligence methods include state machines, decision trees, Bayesian networks, etc. With the development of deep learning and augmented learning technologies, decision-making on complex working conditions and online optimization learning have been realized. Since there are many factors affecting the planning of driving paths in the actual road, it is bound to occupy more computing resources. In order to improve the computational efficiency, Japanese researchers have proposed the research idea of "safety field", that is, to form a typical traffic scene as the input of the deep learning neural network to improve the decision-making efficiency of autonomous vehicles and improve the path planning ability.
Figure 3 Unstructured road detection framework based on machine learning
Intelligent networking
Combined with the development of communication technology, the use of real-time communication technology between vehicles and vehicles, cars and roads, cars and people, and vehicles and clouds can provide further support for the three major elements of artificial intelligence technology in the application of automatic driving technology, such as data, calculation and algorithm, and can also provide solutions to the problems faced by collaborative driving of group intelligent driving systems for multi-model and multi-scenario intelligent driving needs. The specific architecture of the vehicle-cloud collaborative automatic driving system based on intelligent network connection is shown in Figure 4 below.
Figure 4 Schematic diagram of the construction scheme of the vehicle-cloud collaborative automatic driving system based on artificial intelligence
The architecture scheme is divided into two parts: ai-based autonomous driving intelligent terminal and automatic driving cloud system based on big data analysis, which together form a vehicle-cloud collaborative integrated automatic driving system that integrates accurate perception of complex environment, intelligent decision-making of traffic and optimization of driving control. In different driving conditions and application scenarios, the vehicle-cloud collaboration technology can realize accurate driving environment perception, intelligent traffic decision-making and optimization of driving action control, and realize the information and data interaction and collaboration between the vehicle end and the cloud.
The vehicle-cloud collaboration technology of the autonomous driving system based on intelligent network connection mainly solves the problem of insufficient multi-source heterogeneous data fusion and insufficient computing power of front-end equipment, that is, the sampling data of the body sensor node (such as GPS/INS data, millimeter wave radar data) and multimedia data (such as camera images) are transmitted to the cloud database at a certain frequency, and online processing, offline processing, traceability processing and complex data analysis are carried out at the same time. And based on the intelligent driving control model of artificial intelligence integrated application algorithm, it provides a reliable and efficient collaborative control scheme for vehicle decision-making.
The cloud platform of artificial intelligence algorithm application technology is the core part of the automatic driving cloud system, which combines machine learning, data mining and other related technologies to analyze the perception and fusion information and provide a decision-making basis for vehicle control planning. And the use of virtualization technology and network technology to integrate large-scale scalable computing resources such as computing, storage, data, applications and other distributed computing resources to complete the learning and training of artificial intelligence model algorithms, to achieve the training of artificial intelligence models in the cloud, and through the vehicle-cloud collaboration technology to deploy them to the embedded platform, so that artificial intelligence algorithms can be deeply applied in the automotive autonomous driving system.
At present, the application of network technology in the field of automatic driving is mainly concentrated in information services and top-level monitoring, and the realization of high automatic driving through the technical route of intelligent network connection still needs to solve thorny problems such as information security, transmission delay, and network coverage in order to truly land applications.
Intelligent computing platform
Autonomous vehicles are gradually transforming from means of transportation to new mobile intelligent terminals. The changing functions and attributes of the car have led to changes in its electrical and electronic architecture, which in turn requires stronger computing, data storage and communication capabilities as a foundation, and the in-vehicle intelligent computing platform is an important solution to meet these requirements.
The main function of the vehicle intelligent computing platform is to use the environment perception data, navigation positioning information, vehicle real-time data, cloud intelligent computing platform data and other V2X interaction data as input, based on the core control algorithms such as environmental perception positioning, intelligent planning decision-making and vehicle motion control, output drive, transmission, steering and braking and other control instructions, realize the automatic control of the vehicle, and output data to the cloud intelligent computing platform and V2X equipment, and also be able to perform control instructions through the human-computer interactive interface. Human-computer interaction of vehicle driving information.
Reprinted from the forefront of intelligent driving, the views in the text are only for sharing and exchange, do not represent the position of this number, such as copyright and other issues, please inform, we will deal with it in a timely manner.