Lei Feng Network: This article is from the Hard Innovation Open Class | What is the difficulty of ADAS technology? The content is organized from the speech transcript of Chen Mao, founder of Chuanglai Technology.
Chen Mao, Founder of Chuanglai Technology
Guest Speaker: Chen Mao graduated from Huazhong University of Science and Technology in 1990, engaged in GPS navigation software, image processing, machine vision, and artificial intelligence technology research and development for nearly 20 years. He previously founded two GPS application companies, Guangzhou Xutai Communication and Guangzhou Kaimai Navigation. In 2008, he exported independently developed navigation software abroad. In 2015, he won the first prize at the first Maker China Competition held by the Ministry of Industry and Information Technology with his independently developed “Lane Departure Warning System”. He is currently a member of the ADAS Expert Advisory Group of the United States Gree Group.
▌1. What is ADAS?
ADAS is a foreign term, which stands for Advanced Driving Assistant System. It is a technology that actively prevents and avoids driving accidents, rather than passive safety technologies such as seat belts, bumpers, and airbags that are used to mitigate the damage after an accident occurs.
Currently, the concepts of intelligent driving, autonomous driving, and driverless driving are quite mixed and vague. Here, I would like to share my own views. I believe that the initial stage of ADAS is intelligent driving, while autonomous driving is its advanced stage, and the ultimate goal is driverless driving. In the initial stage, it mainly focuses on warning-type systems, which are single-machine driving assistance systems.
In the advanced stage of autonomous driving, it mainly focuses on control-type systems, integrating more with vehicle networking. The ultimate stage of driverless driving relies more on high-precision maps and high-precision GPS. At this point, ADAS is merely an auxiliary driving technology.
Although the concept of ADAS is currently trending towards generalization, the industry generally agrees that only auxiliary driving technologies based on machine vision, image recognition (including visible light images, radar images, infrared images, etc.), and artificial intelligence technology are considered true ADAS technologies. For example, lane departure warning, blind spot detection, pedestrian collision warning, lane keeping, and forward collision avoidance, etc. In general terms, driving assistance functions such as Anti-lock Braking System (ABS), Electronic Stability Program (ESP), navigation, reversing radar, and driving recorders cannot be classified as advanced driving assistance systems.
▌2. What is the Essence of ADAS?
The essence of ADAS is artificial intelligence, and the core is algorithms. In its four stages of perception, cognition, decision-making, and execution, the first three stages involve a large number of algorithms. The perception part mainly consists of sensors, which are responsible for capturing images (visible light, infrared, radar), similar to human eyes. The human eye can automatically adjust the size of the pupil based on the surrounding lighting environment and adjust the focus according to the distance of objects to achieve clear imaging, but similar adjustments in image sensors can only be achieved through algorithms.
The cognition stage is the key and also the difficulty of ADAS. Analyzing whether the observed image contains lanes, vehicles, and pedestrians is not a simple task. Achieving a certain level of accuracy and recognition rate may not be difficult; although the human eye can easily achieve a recognition rate of 99.9999%, achieving a recognition rate and reliability of 99.9%, 99.99%, or 99.999% with machines is indeed a long-term endeavor.
This is because human understanding of image recognition theories is still in its infancy, and even current deep learning technologies are black box technologies, with only superficial understanding of their inner workings.
▌3. Why is ADAS a High Barrier and Long Cycle Industry?
Why is ADAS considered a high barrier and long cycle industry? I believe the biggest barrier to current ADAS technology lies in the recognition algorithms. Whether it is a startup company that has been operating for two to three years, a state-owned manufacturer that has been in business for five to six years, or a university team that has been researching ADAS for over ten years, they all lack sufficient mileage in practical road tests. Without enough road testing mileage and a large amount of data, this directly affects the reliability and robustness of system algorithms.
Currently, the global benchmark in the ADAS field, Mobileye, dominates the market due to its algorithm reliability and precision far exceeding that of other companies. They spent 15 years to achieve a recognition rate of 99.99%. Their first place is indeed well-deserved. The saying “It takes fifteen years to sharpen a sword” also indirectly proves the high barriers of this industry.
Some may argue that a precision rate of 99% is sufficient. If compared to a 100-point exam, it is indeed a high score. However, for applications related to human life safety, even a very small probability event can become a significant probability when multiplied by a large driving time or mileage base. Once it happens, it may endanger lives.
Therefore, the reliability of driving safety has no upper limit, only a lower limit. ADAS products and technologies with a precision rate below 99% are difficult to provide users with a good experience. Currently, the so-called ADAS products on Taobao crowdfunding (which do not even reach a precision rate of 70%) give users a poor experience.
This is because, unlike navigation devices, driving recorders, radar detectors, GPS positioning trackers, and other products that ordinary users can install themselves, we believe that true ADAS will ultimately go down the path of the front-mounted market (front-mounted refers to installation in automotive manufacturing plants).
This is because ADAS requires strict calibration of products to achieve the best warning effect, and professional engineers are needed for installation. If it enters the aftermarket, accurate installation will be a significant challenge. So far, no electronic products that require calibration (i.e., setting parameters related to the installation of camera sensors to achieve accurate warning effects) have appeared in the automotive aftermarket.
ADAS is a technology that directly relates to vehicle and driver safety, concerning personal safety. Manufacturers will inevitably be very cautious when choosing solutions. The environment of vehicles is different from ordinary electronic products; products need to meet automotive specifications, and manufacturers have high requirements for the reliability, stability, and precision of such products.
Additionally, ADAS involves hardware, and the cycle is much longer than that of app projects. On one hand, the iteration cycle of hardware is long; a mature product must undergo multiple hardware iterations, and the development cycle of a new car from a manufacturer is also relatively long, which determines that its projects have long cycles.
▌4. How Can Local ADAS Rise?
Currently, the most seen, heard, and used are foreign ADAS products. So how can local ADAS rise? I believe the market opportunities mainly lie in the following two aspects:
Breakthrough in niche areas. ADAS is a relatively broad field, and it is unlikely for a company to cover everything. Choosing a niche area where one excels as a breakthrough point and focusing limited resources on a specific field can be a viable option to strive to catch up with or surpass industry benchmarks in certain functionalities.
Pursuing integration with autonomous driving. The future trend of the automotive industry is intelligent driving – autonomous driving – driverless driving. If it only stays at the level of auxiliary warning functions, it may be passive in future developments. Therefore, when developing ADAS solutions, it is essential to consider the integration with autonomous driving. Only by doing so can one ensure a place in the industry when the era of autonomous driving arrives.
▌5. Development Trends of ADAS
I hold a very optimistic attitude towards the development trends and future of ADAS.
The development of ADAS will certainly start from offline single-vehicle intelligence, gradually moving towards online ADAS. In this process, it will integrate vehicle networking (V2V/V2I) technology, which, due to the existence of two perception systems (the vehicle’s own sensors and online network sensors), will greatly enhance the reliability and precision of the system.
Moreover, deep learning is essentially a method of machine learning, primarily used for detecting and recognizing objects with indistinct or no features, and is a particularly fashionable method of machine learning in recent years. Currently, the high cost is a barrier to the application of deep learning. Furthermore, in the recognition of rigid objects with more distinct features, deep learning may not necessarily outperform traditional machine learning.
However, in the future, the penetration of deep learning methods into this field will be a trend because, in terms of features, it is difficult or impossible to describe the image recognition field mathematically, such as pedestrian recognition.
But with the outstanding performance of deep learning and the rapid development of computing chip technology and cost reduction, the entry of deep learning into practical application fields will become possible.
The above table introduces the working principles, advantages, and disadvantages of different types of sensors. Through comparison, we can see that different sensors have their own advantages and disadvantages, and each sensor has its applicable range; each type of sensor has its weaknesses.
Therefore, in the future, to achieve all-weather and high-reliability applications, the fusion of multiple sensors will be inevitable.
▌Wonderful Q&A:
Q: ADAS includes many sub-systems, including adaptive cruise control, lane departure warning systems, lane keeping systems, collision avoidance or pre-collision systems, night vision systems, adaptive lighting control, etc. Among these sub-systems, which ones are technically mature? Which ones have the most opportunity for large-scale popularization (applicable to vehicles priced in tens of thousands)? What are the obstacles to large-scale popularization?
A: The ADAS technology involving artificial intelligence is not yet fully mature domestically and is still mostly in testing. Currently, domestic night vision systems basically do not have pedestrian recognition capabilities, so domestic night vision systems cannot be classified as ADAS products.
However, there is a special function, the 360-degree surround view system technology is becoming increasingly mature and is expected to be first popularized in domestic A-class cars. But this system does not actually include artificial intelligence technology, so it is not strictly ADAS. The reason it cannot be widely popularized is that domestic ADAS products still need time to mature, while foreign products are too expensive, so we can only wait.
Q: Does policy have a significant impact on intelligent driving? Will it be like the new energy vehicle industry, where the state stipulates that vehicles equipped with automatic braking (or other systems) must meet certain standards by a certain deadline? What are the main aspects of policy influence on this industry?
A: In fact, ADAS has already begun to appear in a mandatory way through policies. In early March this year, the new motor vehicle safety operation standards clearly stated that buses longer than 11 meters need to be equipped with lane departure warning (LDW) and forward collision warning (FCW) systems.
However, the specific implementation time has not been determined, but it is predicted that it will be implemented in 2017, at the latest by 2018.
