Toyota embraces the pure visual autonomous driving route, what industrial chain signal is released?

Editor's Note: Core Intelligent Driving - A collection of the views of famous experts in production, education and research, a comprehensive analysis of AI chips, third-generation semiconductors and other opportunities and challenges in the era of "big deformation" of automobiles!

According to Jiwei Network, Toyota Motor's subsidiary Woven Planet said in early April that it would use a single visual solution to develop autonomous driving in assisted driving and higher-level autonomous driving projects. That is, in the absence of expensive sensors such as lidar, data is collected through relatively low-cost cameras to promote autonomous driving technology. This means that the veteran car company, which is known for its lean production method, has joined the camp of "pure vision".

Why does Toyota embrace "pure vision"? What industry chain signals does this release? In the PK race of the two major routes of pure vision and lidar, it seems that it has entered the next stage of mass production race from the technical dispute.

Is "pure vision" reliable?

At present, the autonomous driving technology for automotive applications is mainly divided into two camps: one is the lidar faction, which advocates lidar as the leading, with millimeter-wave radar, cameras, etc., to achieve multi-sensor fusion and improve the safety of automatic driving; one is the pure visual faction, which tends to use low-cost cameras, supplemented by artificial intelligence algorithms, to reduce costs.

Ideal CEO Li Xiang once said, "At present, the combination of camera + millimeter-wave radar is like a frog's eyes, which is good for dynamic objects and almost incompetent for non-standard static objects." The progress of vision at this level has almost stagnated, even if it is dynamic, the recognition rate outside the vehicle is less than 80%, and it should not be used as automatic driving. ”

But Tesla CEO Musk insists on the first principle - human driving relies on vision, so neural networks can process visual input and understand the depth and speed of surrounding objects. Tesla firmly believes that with a universal vision system, the vehicle no longer needs any additional information. Last year, Tesla took pure vision to the extreme, removing millimeter-wave radar from its previous autopilot solution in new production models.

"Indeed, from the first principle, there is no lidar when people drive, and we also believe that for high-level automatic driving, vision is one of the most important sensors, and we also believe that the vision scheme can reach a more mature state in the future." Lu Peng, senior director of product planning and marketing of horizon intelligent driving product line, pointed out to Jiwei Network, "But it should be pointed out that not every car company has such a talent as Tesla. ”

Judging from the current low-level autonomous driving schemes around L2, most of them are now a camera plus a millimeter-wave radar or several millimeter-wave radars. It is understood that the mass production of L2-level ADAS solutions that can achieve pure vision in the world is currently mainly Mobileye and Horizon. Lü Peng explained that when visual technology is not mature enough, the input of one more information dimension is to help make up for the shortcomings. But when it's mature enough, it faces another problem: when you get a result from vision, and you get a result from millimeter wave or lidar, and the two are different, which one should you believe? Sometimes there may be intrusive false positives.

"From the now relatively mature low-level automatic driving to the high-level automatic driving, in fact, it is also such a process." Lü Peng said that in general, from the perspective of the current industrial chain, he believes that the coexistence of pure vision and lidar routes will still be relatively long, but in the future, as the technology becomes more and more perfect, the proportion of vision in the automatic driving program will gradually increase, and the pure vision program may be a final form that is more in line with the cost and first principles in the future.

Dr. Ming Liu, a senior member of the IEEE, director of the Center for Intelligent Driving at the Hong Kong University of Science and Technology, and director of the field of robotics and autonomous systems, pointed out that machine vision can theoretically achieve full autonomous driving functions. He explained that the core of the visual function of autonomous driving is to provide its own positioning and how to understand the surrounding environment, so as to achieve navigation and obstacle avoidance, as well as human-computer interaction and other functions. With the development of deep learning and reinforcement learning, machine vision has achieved breakthroughs from simple two-dimensional object recognition to three-dimensional reconstruction, semantic recognition, and enhanced iterative obstacle avoidance functions, coupled with the integration of big data parallel processing systems such as cloud computing and V2X networks and on-board autonomous driving systems, which can well achieve full automatic driving.

However, to achieve a truly mature "pure vision" technology, so as to achieve the landing of fully autonomous driving, a lot of engineering practice is needed. In the current mainstream visual solution, the vehicle camera mainly includes the internal view camera, rear view camera, front camera, side view camera, surround view camera and so on. Dr. Liu Ming analyzed that the current visual scheme of autonomous driving cameras relies heavily on a large number of high-definition photos to fit deep learning models, thereby improving the accuracy of recognition and positioning. The emergence of different degrees of lighting and unknown objects increases the difficulty of machine vision recognition and positioning of simple cameras, and from an engineering point of view, a series of issues including noise reduction, V2X network connection and the use of cloud computing to assist large-scale intelligent transportation need to be further explored. To achieve fully autonomous driving, this needs to be supplemented by a lot of engineering practice.

It is reported that Tesla is building a deep learning system for the training of the autopilot system, in addition to processing depth, speed and acceleration information, but also at the same time to detect the target. This requires a huge dataset of millions of videos to train a deep learning architecture, and creating datasets for self-driving cars is tricky, and engineers must ensure that the datasets have diverse road settings and infrequent edge situations.

The industrial logic of the next step of advanced autonomous driving competition: rapid mass production

The industrial logic behind Toyota's embrace of "pure vision" is actually not difficult to understand.

The first is for cost control considerations. From the perspective of model level, Toyota's models cover a large number of mainstream users who consume mid-range models, and they are of course more demanding than Tesla in terms of cost control. Although the price of lidar is obvious, it is still a lot more expensive than the camera.

Secondly, Toyota does not need to use lidar to support the brand's high-end appeal. Some industry views believe that new forces favor lidar, in addition to the advantages of more accurate perception and positioning, there is also a key reason that lidar comes with a sense of high-end. At the same time, in comparison, the new force car companies have no historical baggage, conform to the trend of intelligence, cut from high-end cars, it is easier to carry out product layout from top to bottom, and lidar, although a little expensive, is acceptable for high-end models with higher premiums.

In this regard, Lu Peng pointed out that in terms of high-end automatic driving, car companies are currently facing the biggest challenge, one is the electrification transformation, and the other is that many domestic cars are doing "brand up". From the perspective of brand upwards, the sooner this high-level autonomous driving function can be launched, while having a better experience and controllable costs, it is very critical, and these are actually challenges for car companies.

"Now the key to competition is mainly to accelerate the commercialization of landing." Lü Peng pointed out that at present, most car companies will tend to choose a program with lidar, mainly to do a high-level automatic driving of a pure visual solution, to the mass production landing, and the improvement of its program needs to take longer to achieve.

For Toyota, in the high-level automatic driving landing competition, in addition to the cost factor of choosing a pure visual route, how to quickly land is bound to be a key consideration. All along, in the intelligent and electrification transformation, Toyota is not a leader, and its transformation strategy has always been very conservative. However, at the end of last year, Toyota changed its conservative posture and launched more than ten pure electric models in one go, releasing a signal to catch up.

However, Toyota as a whole has been in a state of obscurity for a long time in terms of investment and research in the field of autonomous driving. Toyota set up an autonomous driving department in 2018, but did not disclose too much about the degree of self-research on self-driving. Since then, Toyota has also invested in ATG, the autonomous driving division spun off by Uber, Momenta, a Chinese intelligent driving solution company, and Xiaoma Zhixing, an autonomous driving startup, to develop autonomous vehicle technology, but it has not yet seen a major breakthrough. Perhaps, for Toyota, crossing the complex process of multi-sensor fusion and directly using pure vision as a breakthrough is a "shortcut"?

This may be a glimpse of what Toyota's Woven Planet has said – "It's not enough to collect a small amount of data from a small batch of very expensive self-driving cars, we need a lot of data." Woven Planet believes that equipping vehicles with expensive sensors is not a sustainable approach, and the large amount of data collected by cameras has great flexibility, which can expand richer capabilities at a certain hardware cost. After all, with toyota's global popularity and penetration, it can grasp a considerable amount of data from it to make its autonomous driving system fully deep learning.

"I think in the end, everyone is fighting for user experience. If a pure visual solution without lidar and a program with lidar can be achieved with lidar, if the two can achieve the same safety and user experience, then the market will naturally be biased in the direction of pure vision, after all, this directly involves the cost of the whole vehicle. It's a step-by-step process. Lü Peng pointed out.

Whether it is pure vision, lidar, or other technical routes, the autonomous driving solution that ultimately wins in the market must be a good balance between cost and safety.

(Proofreading/Jimmy)