In the AI era, autonomous driving technology is evolving rapidly

Source: Automotive Test Network

Author: Beidou

Introduction to Beidou: 10 years of experience in the development and management of smart cockpits and navigation and entertainment systems, 3 years of experience in productization of autonomous driving, and 5 years of experience in building an autonomous driving simulation test environment.

In recent years, with the empowerment of policies and markets, the autonomous driving industry has accelerated its implementation, and the basic supporting industry chain and market development have become increasingly mature. Since 2020, the autonomous driving industry has officially entered the "golden decade". It is expected that by 2030, my country's autonomous vehicle market share is expected to exceed 50%, and the autonomous vehicle service market size is expected to reach 1.3 trillion. From the perspective of technological development trends, my country's autonomous driving technology industry is currently evolving from single-vehicle intelligence to the era of vehicle-road collaboration, and it is AI (artificial intelligence) technology that supports this evolution. The third AI boom caused by deep learning has promoted the arrival of the AI era. **This article focuses on analyzing and introducing the evolutionary promotion of the application of AI technology in the field of autonomous driving in the AI era. **

Figure 1 Mind map of autonomous driving evolution content

Evolution of Autopilot System

1. AI in autonomous driving image analysis

In the automatic driving system, the vehicle is equipped with a variety of perception sensors such as cameras, millimeter-wave radars, and lidars. The system will analyze the data obtained by perception, and make vehicle control judgments based on the results of AI data analysis.

For the raw data obtained by perception sensors such as cameras, the automatic driving system cannot directly judge, because the system lacks the ability to classify things like a baby in the early stage. Therefore, we must first classify and classify the data one by one. This work is data annotation. Classify and mark all traffic and road-related elements of various traffic facilities (lane lines, road signs, traffic lights, etc.), various traffic participants (pedestrians, bicycles, passenger cars, commercial vehicles, special vehicles, etc.).

Figure 2 Data annotation map of real traffic scene

The processing unit of the autonomous driving system will use these annotation and classification results as the basis, and the AI will learn the characteristics of various classified objects. The more basic data, the more prominent the features, and the higher the accuracy of object discrimination. AI is like the brain of the autonomous driving system. It analyzes the characteristics of each object and learns the appearance characteristics and movement habits of the object bit by bit. The AI brain gradually becomes smarter through such repeated learning work. While identifying the category of objects in the image, it can also grasp the overall condition of the object. This is the application in the field of computer vision related technologies that we are familiar with. In addition, it is also possible to automate the classification and labeling work by AI.

2. AI in autonomous driving decision-making and judgment

Through computer vision, the system can realize the overall status of the data obtained by the sensing sensors, and based on this, make judgments and decisions on vehicle control. This is exactly how AI is advancing the evolution of autonomous driving technology.

Based on sensory data, AI will make the same judgment as human driving habits in the shortest possible time. In order to realize the real-time performance of image processing and the instantaneity of judgment and decision-making, there is a strong demand in the field for the development of high-precision AI based on powerful data processing capabilities.

3. AI in automatic driving predictive control

One of the elements of judgment decision is "prediction". How the vehicle or pedestrian driving in front will move next? AI needs to predict the possible actions of all objects in the traffic environment in advance, and implement vehicle control based on the prediction.

The "trolley problem" in the eyes of AI

Suppose the self-driving vehicle is driving on a one-way, one-lane road with trees on both sides, and the brakes suddenly fail. There is an old man on the road in front of him, and a baby crossing the road. We should make a choice. This is actually " Trolley Problem” a deformed scene. In the case of exceeding the scope of the system's predictive ability, autonomous vehicles cannot make decisions and judgments in extreme situations, and the state of decision-making conflict will become the Achilles' heel of system security. Based on common sense judgment logic, in order to avoid endangering the safety of personnel, the only option is to make a sharp turn and crash the self-driving vehicle into a tree. When the system is forced to make the ultimate choice that cannot avoid driving or occupant injury, what decision should the AI make actually reflects part of the developer's intentions, whether it should protect people other than its own car or protect its own car's driver? What about the crew. Or should we judge based on the number of people, or should we do our best to slam on the brakes and let nature take its course.

Figure 3 Cartoon version of the deformed version of the “Trolley Problem”

In fact, this issue has always been controversial, and it is not easy for humans to draw accurate conclusions. However, in some areas, the government has passed legislation to regulate similar issues. For example, the "Automated Driving Law (Amendment of Road Traffic Law)" passed and implemented in Germany stipulates: "When there is an unavoidable risk of personal injury, the accident prevention system should have the decision-making ability to not weight human lives based on personal characteristics." This also provides AI with a clear decision-making direction for such problems.

Evolution of path planning for autonomous driving systems

Comprehensive judgment of the route and destination and planning the most appropriate path are one of the essential skills for autonomous vehicles. When planning a route, it is not only necessary to consider traffic congestion predictions and road construction between destinations, but also to select the most appropriate lane-level route planning, and while ensuring route convenience for multiple passengers, the system must Instantly determine the order in which to execute the path to achieve the most effective and shortest path planning.

In order to continuously improve the system capability, it is necessary to conduct risk analysis on the actual accident rate of the planned route, conduct data analysis based on the road conditions, number of turns, number of signal lights and other information of the roads passing through the planned route, gradually optimize the route planning strategy, and finally improve the system planning capability .

Figure 4 Schematic diagram of multi-way point path planning for autonomous driving

When self-driving taxis are used, multiple vehicles operating in the same area may demand vehicles at the same time. Vehicle dispatching also requires the self-driving system to plan the most appropriate path for all taxis. Moreover, the prediction of when and where vehicle demand will occur is also a basic function that subsequent autonomous driving systems need to implement for vehicle dispatching. Technology that adds predictions of future demands to complex application scenarios and can provide corresponding decision-making results in an instant is currently only capable of using AI.

Evolution of human-computer interaction in automatic driving system

In autonomous vehicles without a driver or safety officer, the most important thing is to accurately grasp the status and needs of passengers. Then the system will replace the driver to complete the reply or report of the current vehicle driving status information, and complete the necessary communication with the passengers during the driving process. These requirements are exactly what AI is good at.

Figure 5 Schematic diagram of interaction between autonomous vehicles and passengers

Communication between passengers and self-driving vehicles will also use speech recognition technology currently widely used on mobile phones and tablets. Although human language often has extended meanings that are more difficult to understand in addition to the surface meaning, due to the intervention of AI, the system's understanding ability will gradually improve from the basic clear instruction "I want to go to the hotel" to the understanding that passengers need to further understand The vague instruction of real needs "I want to eat delicious food" is actually what AI is good at.

In addition to understanding various instructions from passengers, AI can also analyze the status of passengers based on information collected by sensors such as cameras in the car, allowing them to think independently and execute corresponding measures independently. For example, when the AI determines that the occupant is sleeping, it can consider dimming the lights in the car and playing relaxing music to aid sleep. In the case of continuous coughing and elevated body temperature of passengers, proactively remind nearby pharmacies and clinics along the route.

In future self-driving cars, AI will treat passengers like distinguished VIP customers, providing meticulous service. Especially in 2023, the release of ChatGPT will set off another wave of AI boom. In autonomous vehicles, AI answering other people's questions with voice is an inevitable functional evolution.

The evolution of cloud and edge computing in autonomous driving systems

As the data that autonomous vehicles need to process gradually continues to increase, simple vehicle terminals are gradually unable to meet the computing power requirements for data processing. In order to meet the processing requirements, the data is sent to the cloud, and the data is processed and analyzed by the AI in the cloud. The AI analysis results can be sent back to the self-driving vehicle terminal at any time. This data processing method has also become an automatic driving with the development of AI. One of the standard architectures.

Figure 5 Schematic diagram of cloud learning and edge computing prediction

There will definitely be data extension during the entire data transmission process, and the real-time nature of the data may be compromised. To solve this problem, on the one hand, it is necessary to optimize and improve the communication rate and communication data volume of wireless communication. On the other hand, AI has also promoted the development and application of edge computing technology in autonomous driving vehicles, which has deepened the data preprocessing capabilities of the vehicles. evolution.

Summarize

Generally speaking, since entering the digital era, AI has deeply empowered all walks of life. The arrival of the "big model" era has allowed the AI industry to intersect with more traditional industries, including automobiles, and have evolved them. and promotion effect. AI has gradually replaced drivers, improving and evolving safety, accuracy, and comfort in many aspects such as driving environment perception, path planning, vehicle control, and passenger interaction. It is expected that in the AI era, autonomous driving technology will maintain its rapid evolution and promote the comprehensive application of autonomous vehicles.