Basket shares | how to "verify the right body" for autonomous driving safety

Obviously, at the technical level, automatic driving is bringing the goal of "safety first" to reality, and at the policy level, the intervention of the "national team" means that the government recognizes and supports the industry - automatic driving will be safer than human driving, and it must be implemented from the theoretical level to the practical level.

For the safety of automatic driving, we can still comprehensively analyze the "fundamental lifeblood" of automatic driving from the theoretical assumption of "how to ensure safety", to the technical architecture of "ensuring safety", and then to the technical practice of how to truly achieve "safety first".

For Safety: Why Does Autonomous Driving Have to Outperform Human Drivers?

It's not hard to understand this layer of the problem, we just need to look back at the original intention when autonomous driving technology was created.

Unlike most people, who think of providing a simple two-point commute "from point A to point B", the first task in the development of autonomous driving technology is actually to reduce traffic casualties and improve road traffic safety. Through autonomous driving technology, the car can fully identify various dynamic and static targets on the road with the help of a comprehensive perception system, intelligent decision-making system and accurate execution system, and achieve autonomous decision-making and vehicle operation - which means that in a complex and changeable driving environment, autonomous driving has a theoretically higher upper limit than human driving.

According to the research of the "Autonomous Vehicle Traffic Safety White Paper 2021", compared with human driving, self-driving cars can effectively perceive the surrounding environment in advance and can make the right courtesy to other traffic participants. In the scenario of human driving accidents, under the normal driving of autonomous vehicles, it can significantly reduce the risk of conflict between vehicle pedestrians, non-motorized vehicles and other motor vehicles, and further protect the life safety of vulnerable traffic participants. And in the bicycle accident caused by the driver's own reasons, the self-driving car effectively avoids it with strict safety algorithm control and further strengthens its own safety guarantee. Therefore, in the face of China's road traffic environment, the development of automatic driving can greatly improve road traffic safety.

Compared with human driving car accidents mainly caused by their own mistakes, autonomous vehicles do not need or rarely need the driver's operation during the driving process, and the driving of the vehicle will be more standardized, thus reducing the traffic accidents caused by driver factors.

In the face of the cause of the accident caused by the subjective error of the driver, because the self-driving car will not have the influence of fatigue and emotions, perfect decision-making planning can ensure that the vehicle drives in accordance with the traffic regulations. At this time, the driver's subjective errors in the failure to give way according to the regulations, the speed is too fast, the illegal use of the lane, drunk driving, violation of traffic lights and fatigue driving and other phenomena can be almost completely solved, which is equivalent to the normal driving of autonomous vehicles can completely reduce the subjective error of human driving, that is, 80% of the accident causes of human driving can be effectively avoided. Traffic safety has been greatly improved.

In the environment of mixed traffic flow of autonomous vehicles and human driving cars, self-driving cars cannot regulate the driving behavior of other traffic participants, and cannot completely avoid accidents caused by other traffic participants, but autonomous vehicles can perceive and predict risks and dangers in advance, at this time, self-driving cars can effectively alleviate the casualties caused by accidents, and can reduce the collision risk of secondary accidents.

Therefore, with the blessing of automatic driving technology, self-driving cars can effectively avoid various problems that often occur when humans drive, and even one day completely eradicate them, which is completely feasible in theory; but correspondingly, the thinking logic of automatic driving technology is different from that of humans, so today, self-driving cars will always make some "low-level mistakes" that are unimaginable to ordinary people, but these "low-level errors" can be solved through the iteration of technology. For the industry, this requires continuous improvement through perception, cognition, labeling, verification and other means, and for laymen, it is also necessary to continuously increase the understanding of autonomous driving technology and give more understanding.

Ensuring Safety: How Can AI Escort Autonomous Driving Safety?

Through the above description, we know that "safety first" has always been the core concept and value of automatic driving, and its ultimate goal is that the automatic driving system, as an AI driver, can replace humans to achieve complete "unmanned driving" - perhaps there is no 100% safety in this industry, but this does not prevent the continuous development and improvement of automatic driving technology towards the goal of 100% safety. In order to achieve this goal, autonomous driving technology needs to be continuously upgraded at the three root levels of perception, algorithm and computing power, both to be perfect and redundant, in order to finally achieve a complete replacement for humans.

According to the white paper research, the safety technology guarantee for automatic driving can be guaranteed from the following four aspects.

At the perceptual level, environmental perception is a prerequisite for autonomous driving. The environment perception system integrates the advantages of multiple sensors such as lidar, millimeter-wave radar, and cameras to achieve a 360° visual distance around the body, and to stably detect and track the behavior and speed orientation of traffic communicators in a complex and changing traffic environment, providing scene understanding information for the decision-making planning module.

The perception algorithm adopts the framework of multi-sensor fusion or multi-camera fusion, and is based on deep neural networks and massive automatic driving data to more accurately identify the type of obstacle and track the behavior of obstacles stably, providing stable perception capabilities for downstream decision-making modules. Among the two mainstream fusion solutions, the perception system based on the multi-sensor fusion scheme is the most popular choice, which can form redundancy through heterogeneous sensing pathways, providing high fault tolerance for the automatic driving system, thereby improving system safety.

When it comes to the advantages of perceptual systems specifically, we can compare them with humans. Human perception as long as the brain through the eyes to obtain visual information, and the eyes of the distance of vision is only 120 °, not only the central pixel is much higher than the edge pixel, but also by the night, fog and other environmental influences, it is difficult for humans to obtain full-dimensional perception of the environment when driving; and the perception system using multi-sensor fusion scheme can reduce these effects, and through the continuous progress of perception technology, it will eventually be completely offset. This is also why domestic autonomous driving companies have chosen the multi-sensor technology route, which we will talk about below.

At the algorithm level, the algorithm is the brain of autonomous driving technology, and the perception, prediction, decision-making and execution actions based on the algorithm will directly affect the ability and effect of vehicle autonomous driving. The algorithm takes "safety first" as the principle, and based on traffic safety norms and consensus rules, can plan a safe, efficient and comfortable driving path and trajectory for the vehicle - this is also a forward-looking function, only in the case of the vehicle turning left through the traffic light, the mature algorithm system can be reasonable planning and control for the vehicle's speed, driving trajectory, steering angle, etc., to ensure that the vehicle can be naturally integrated into the same direction of traffic flow after the left turn under the premise of making the passenger feel comfortable with the speed and rollover. And to ensure that there is no pressure line, no accidents.

In order to achieve this idealized function, we must improve the generalization ability of the algorithm, specifically through the application of data mining and deep learning algorithms to achieve intelligent planning of driving behavior. Among them, there are deep learning models trained based on massive test data in conventional scenarios to ensure the safe, efficient and smooth passage of autonomous vehicles in conventional driving scenarios, and there is also a set of safety algorithm layers that design a series of safe driving strategies for various typical dangerous scenarios to ensure that autonomous vehicles can make safe driving behaviors in any scenario.

With more and more road test data for autonomous driving, accumulating a large amount of data on extreme scenarios, the core algorithm of autonomous driving continues to evolve into an "old driver" who predicts in advance and drives safely and cautiously through a data-driven deep learning algorithm model.

At the level of computing power, in order to serve the continuous improvement of the generalization capability of the algorithm, the computing power of the chip equipped with the terminal must also be continuously upgraded. At present, the market generally believes that only more redundancy in computing power can provide more space for the subsequent needs of "hardware embedding and OTA upgrade", but at the same time, in the context of the concept of "software-defined car" becoming the industry consensus, the demand for SoC chips with higher computing power, higher throughput, lower latency and low power consumption is becoming more and more urgent - in this regard, some media have pointed out that the co-evolution of chips and algorithms should not be bundled, but chips give algorithms more freedom. Free up the shackles of algorithm implementation. If the chip itself has insufficient computing power and low openness, the synergy between the chip and the algorithm can only play a role in cutting enough. This means that what is really needed for autonomous driving technology should be based on large computing power and high compatibility, while providing an open and flexible development environment to minimize the time and cost of algorithmic adaptation of chips.

It is obvious that with the mass production of more and more car companies with high-end driving assistance systems, the amount of data on automatic driving will undergo structural changes, and the change is not only in the exponential expansion of the data volume itself, due to the transformation of data types, automatic driving technology will pay more attention to the bandwidth of data transmission and the efficiency of data handling. Therefore, with 5G + V2X technology, the establishment of a cloud supercomputing center will also become a key step at the level of computing power, and this kind of "open plug-in" measure can not only provide computing power support for deep learning models trained on massive test data, but also ensure efficient collaboration between the cloud and the terminal.

In short, perception, algorithm and computing power are actually the knowledge, experience and logic summarized by human driving in the past hundred years, translated into the language that the machine can understand to allow the machine to learn itself, on the basis of allowing the machine to master the logic of human driving, and then with the help of stronger perception ability than humans, faster reaction thinking speed, more reference cases and strict compliance with various laws and regulations, so that automatic driving can reach and surpass human driving and become safe enough. With the help of the large computing power platform and supercomputing center, automatic driving technology has been able to calculate the demand for power, and what the major autonomous driving companies should do next should be to improve the number and quality of multiple sensors to provide comprehensive and even redundant perception data, and then use algorithms to efficiently absorb and digest, so as to promote the safety of automatic driving technology.

Of course, in this process, the more data collected should be as good as possible, and the corresponding cost should be as low as possible, so as to ensure that the competitiveness of an autonomous driving company is at the forefront of the industry.

Be safe: How do leaders take safety to the end?

Under the efforts of major car companies and autonomous driving companies, automatic driving technology has gone through the first half, and the head autonomous driving enterprises represented by Tesla, Mimo Zhixing, and Xiaopeng Automobile have realized advanced driving assistance functions in high-speed scenarios, and urban scenes will also be launched this year, and will completely realize the full-scene function at the end of the year, laying the foundation for true automatic driving. And traditional car companies are also going all out to invest in the wave of intelligent automobile technology.

As mentioned at the beginning, Mercedes-Benz is becoming the traditional car company that has gone the farthest on the autonomous driving route - after obtaining the L3 level international certification, Mercedes-Benz recently directly publicly stated that when the Mercedes-Benz car driver equipped with Drive Pilot opens the vehicle's advanced driver assistance system, they are no longer legally responsible for the operation of the car. In the event of a car accident, Mercedes-Benz will bear the relevant responsibility.

Mercedes-Benz's passenger car driver assistance system is called The Intelligent Pilot System (Drive Pilot), which has reached L3 level, which adopts camera + lidar + high-precision map technology route, and can open the L3 automatic driving function on specific highways supported by high-precision maps, but the speed is required not to exceed 60km/h. It complies with the first L3 level regulation ALKS that came into effect at the beginning of this year, although this makes Mercedes-Benz the first car manufacturer in the world to obtain an "L3 level pass", but due to the immature ALKS regulations, its hard requirements for L3 level are very conservative, so Drive Pilot can not yet support high-level auxiliary driving functions such as lane change, in and out ramp.

Compared with traditional car companies, the achievements of new car-making forces in the field of automatic driving are very remarkable, and the most controversial, but also the most effective and promising representative is Tesla. Tesla adopts a pure visual technology scheme of sword-and-blade style, the visual information is first identified, analyzed and calculated by the FSD chip by the camera by the AI neural network technology, and part of the computing power is borne by the HW 3.0 chip at the end of the car, and the computing power exceeding the computing power of the car is apportioned to the Dojo supercomputer.

Tesla explains it as, "We want to be able to build a neural network connection similar to the visual cortex of animals, simulating the process of brain information input and output." Just like light enters the retina, we want to simulate this process through the camera" - according to this logic, Tesla used neural network architectures such as HydroNet and transformer at the algorithmic level for deep data learning, and finally used this as an advantage to expand the 2D images acquired by pure visual solutions into a 4D vector space with timing and air order.

The perception system based on the multi-sensor fusion scheme can form redundancy through heterogeneous sensing pathways, provide high fault tolerance for the automatic driving system and thus improve system safety, which is the understanding of the new domestic car-making forces represented by Xiaopeng Motors that are different from Tesla. In order to adapt to the more complex and changeable traffic environment in China, Xiaopeng Automobile adopts camera + lidar + high-precision map technology route, with high-precision map as the leading, visual system as an auxiliary, lidar as redundant support, and 30 TOP NVIDIA Xavier support, which can not only achieve centimeter-level urban positioning capabilities, but also connect different scenarios such as parking lots, highways, and urban roads.

Also using the technical route of camera + lidar + high-precision map, the self-driving company Behind Great Wall Motors has also summarized a set of forward-looking methodologies and made great strides on the road to success. Taking HPilot 3.0, which will be landed this year, as an example, at the computing power level, The Wisdom Act provides a self-developed domain controller "Little Magic Box 3.0" with a veneer computing power of up to 360TOP, while at the perception level, it relies on a high-precision map and the first echelon of hardware configuration composed of 14 cameras, 5 millimeter-wave radars and 2 lidars to achieve assisted driving in the high-speed domain and the urban domain that will be launched soon.

At the data level, Zhixing believes that data is the core of autonomous driving technology. After analyzing and summarizing the accumulated mileage data of massive users, the development curve of autonomous driving capability is obtained: F=Z+M(X). Among them, F stands for product power, Z stands for the first generation of products, and M is a function that transforms data into knowledge, including data acquisition, expression, storage, transmission, calculation, verification, and the impact on cost and speed, that is, data intelligence system MANA. Guided by MANA, and with the help of massive data brought by Great Wall Motor's huge mass production advantages, the algorithm is trained to collect, test and verify on a large scale.

At the algorithm level, in order to achieve efficient fusion of multi-sensor sensing data, MANA introduces the transformer multi-layer neural architecture to do pre-fusion in space and time, first transformer encodes image features and decodes them into three-dimensional space, and coordinate system transformations have been embedded in the calculation process of self-attention to achieve pre-spatial fusion; secondly, time series data, as the old line of Transformer, can be naturally extracted to time series features.

Transformer solves the problems that MANA faces at the perceptual level, while at the cognitive level, MANA uses CSS to solve the safety of autonomous driving. The core of the self-developed safety cognitive model CSS is not limited to ensuring that the autonomous driving system is not only limited to ensuring that it does not actively make mistakes from a purely mechanical point of view, but also fully considers the understanding of the behavior of other traffic participants learned from the data and the historical experience of time-lapse; and above the safety bottom line, MANA can learn comfortable and more efficient quantitative standards from the data, so that the autonomous driving algorithm can better handle the complex driving scenarios and formulate driving strategies that are more in line with user preferences. And through the automation of scene mining, reinforcement learning, simulation engine to build a cognitive intelligence closed-loop system, continue to extract knowledge from massive human driving data, and quickly iterate the ability of vehicle-side cognitive algorithms.

From Mercedes-Benz to Tesla to Miller Wisdom, we see that the self-driving technology is gradually outlined, its functions have not only become more powerful, but also more and more secure, which is exactly the winning advantage brought about by mana's forward-looking strategic planning by Mana, that is, taking data as the core and obtaining data more efficiently & at a lower cost. With data as the guarantee and computing power algorithm as the support, automatic driving technology will become more and more intelligent, and naturally it will become more and more secure.

In 2022, safe autonomous driving is coming

Can autonomous driving really ensure safety and be safe? I believe that seeing this, you have a clearer understanding and have established a preliminary confidence in the future of autonomous driving technology.

With the support of national policies, autonomous driving technology is no longer a castle in the air. The application of data intelligence in autonomous driving has become a technological accelerator to promote the safety of autonomous driving. Technological iteration plays a positive role in positive empowerment for the landing of policies and the introduction of relevant standards.

Entering 2022, major head autonomous driving companies at home and abroad have entered the high-level assisted driving city scene, and in turn are important stages, with the second half of the year as the key node to achieve full-scene high-order assisted driving, and then achieve full automatic driving. In China, the autonomous driving enterprises represented by The Wisdom of the End have planned a clear blueprint for automatic driving technology, and the national policy has also affirmed and supported it, and automatic driving will gradually become the "new normal" in the daily travel scenarios of the public in the next few years.

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.