As we all know, there are a wide range of technical differences in the field of autonomous driving technology perception. The multi-sensor fusion group based on laser radar and the visual priority of camera-based technology have carried out long-lasting technical cooperation. Recently, the flag-bearer of the lidar camp, Waymo, and the leader of the camera camp, Tesla, once again confronted each other.
The new round of confrontation dates back to April 22nd. On Tesla Autopilot Investor Day, Elon Musk fired again at Lidar.
Lidars are all fools, and any company that relies on lidar is destined to fail. They are expensive, unnecessary sensors.
Why are you so arrogant?
Elon has always been unscrupulous and arrogant. This personality is closely related to his success in entering a new field for the past 20 years.
On the investor day of Tesla’s launch of autonomous driving chips, analysts questioned Tesla’s technical route for the Nth time. Considering that Elon has completely taken over from the end of 2015, directly leading and recruiting the Autopilot team is not a question of the potential of the camera, which is questioning Elon’s judgment on the technology landscape. Elon, which has been explained N times repeatedly, has finally erupted.
Elon broke out this time, and many practitioners of the lidar camp responded. However, Waymo, which is most anticipated by the people of the melon, did not speak at the time.
Waymo is waiting for the opportunity. At the Google I/O 2019 conference on May 8th, Waymo CTO and vice president of engineering Dmitri Dolgov and Waymo chief scientist Drago Anguelov launched a full counterattack.
Elon's use of a camera alone to remove the lidar is "very risky."
You can imagine autopilots only through the camera, but you need the best camera system to solve the problem. So this is a very big bet, you can implement it, but it is very risky and unnecessary.
Anguelov believes that Lidar helps Waymo create a safer user experience for users.
We have more data and more accurate, and we are more likely to build the right simulation environment. Lidar helps cars determine how cars and other objects on the road interact. If you only use a camera, all of this is much harder and has more limitations.
Dolgov’s response made Tesla completely in a dilemma.
(For cameras and lidars) We are not one or the other, we have both. Everything is done to fully capture the world of two perception systems and combine them in an understandable way to have the most powerful and secure system.
Lidar is essentially nothing expensive, and we have significantly reduced the price of laser radar from the first generation to the current. You can imagine how much space will cost down as we expand.
We have a road map and some plans for expansion (out of the autopilot fleet) outside of Phoenix.
(In December 2018, Waymo officially launched the Waymo One, a self-driving car taxi service in a limited area in Phoenix, Arizona, USA.)
For Waymo’s rebuttal, Elon lost patience and responded to Twitter. “Anyone who purchases Tesla’s fully automated driving function can use Autopilot this year (later) without a manual intervention from California to New York. ”
Looking at these world-class scientists and technologists in the field of autonomous driving, I feel very lost, so it is not good. Self-driving cars are a new thing. We should let ordinary consumers understand why the differences between Waymo and Tesla are of little value. Why the perception of autopilot different technical routes will ultimately be the same.
Let’s start with Waymo’s counterattack. It has to be said that regardless of how Elon responds, Waymo’s precision strikes are crucial.
This creates a perception in the minds of consumers: camera = danger; lidar = safety. In fact, the opposite of Tesla is not Waymo alone, how big is the Waymo camp? Foreign media The Verge’s title is in the phrase.
IT’S ELON MUSK VS. EVERYONE ELSE IN THE RACE FOR FULLY DRIVERLESS CARS
The Tesla CEO is forging his own path toward full autonomy
Let’s start the discussion one by one.
Elon said that the main problem of laser radar is “expensive” and “unnecessary”. Lidar is safer and has solved the “unnecessary” problem. So “expensive”?
Unlike most automakers and autopilot companies, Waymo has long since squandered the product of Velodyne, the top manufacturer in the field of laser radar, but has taken the path of independent research and development. Similar to the Tesla self-developed chip, the independent development of laser radar brings unique competitiveness to Waymo.
For example, the equivalent performance of the laser radar, Waymo self-research costs far less than the purchase of Velodyne. As early as the beginning of 2017 at the Detroit Auto Show, Waymo CEO John Krafcik announced that Waymo had reduced the cost of lidar by more than 90%.
Waymo’s previous purchase of Velodyne’s 64-line laser radar cost around $75,000, a drop of more than 90% means that Waymo is keeping costs below $7,500.
Considering that Autopilot 2.+’s BOM cost control is around $2,500, Elon can still maintain its existing attitude: Waymo’s work is fruitful, but today’s laser radar cost is far from commercialization.
But don’t forget that Waymo’s plunging more than 90% is in 2017, and the Waymo Laser Radar team has been working hard for performance improvements and cost reductions.
On April 24th, Forbes exposed the next generation of sensing systems that Waymo is testing, including laser radar systems and camera systems. Waymo did not disclose details of improvements or improvements to the new lidar, but confirmed that “new sensor systems have been testing in the San Francisco Bay Area in recent weeks.”
If we look back at the various developments that Waymo has made over the past two years, we will find that the speed of Waymo commercialization is still quite impressive.
In addition to the above progress, another reason that Waymo deserves attention is the picture below.
In addition to the laser radar on the roof, Waymo is also in front of and behind the car, with short-range, high-resolution mid-range and high-performance long-range laser radar on the left and right sides. The Waymo stacking madness comes from the control of the cost of the self-developed sensor, and the full sensor arrangement in turn achieves the undisputed first in the field of automatic driving at the technical level.
How do we view Waymo, Waymo is working on the road to commercialization of auto-driving cars without the crowdsourcing team collecting huge amounts of data, the high cost of laser radar, and the lack of improved algorithms.
General Cruise is catching up quickly and growing into a dazzling new star. However, if we comprehensively judge from the perspectives of technology, cost, and commercialization, we should not talk about Tesla, which does not take the usual path. The automatic driving field should be divided into Waymo and others.
However, the real world is not allowed to throw away Tesla. Let’s talk about this company. Go back to the beginning of the article, you will observe the interesting details.
Waymo Chief Scientist Drago Anguelov criticized Tesla for using the camera “very, very risky and unnecessary”, but regarding the automatic driving with the camera landing, his evaluation is “you need the best camera system to solve the problem”, “You can Realizing it is a huge bet and “all of this is much harder.”
As a Ph.D. in computer science, artificial intelligence, and machine learning at Stanford University, and a lead scientist at Waymo, Drago Anguelov has a clear understanding of the potential of camera perception and the boundaries of computer vision + artificial intelligence. The challenge of auto-driving based entirely on the camera is huge, but it doesn’t mean it’s completely infeasible.
Why is Waymo’s perception focused on laser radar? This is determined by historical factors. In 2009, when the Google Unmanned Vehicles project started, artificial intelligence technology has not made significant progress. This directly leads to the visual perception of a self-driving car based on the camera at that time is completely infeasible.
Who gave Elon the courage to stand in the opposite of the industry?
In 2012, the depth of the ImageNet Large-Scale Visual Recognition Challenge driven by the Deep Convolutional Neural Network dropped dramatically to 16%, which is considered an unprecedented breakthrough and has become the starting point for a new wave of artificial intelligence. This directly shakes the rapid development of artificial intelligence and computer vision recognition technology.
So today, Waymo’s perception system and Tesla’s perception system are obviously different.
Waymo is based on lidar, and point-cloud based relies on data from points (or distances measured to objects) in 3D space collected by active sensors.
Algorithms may involve deriving structures from a large number of points by point density, geometry or pattern to detect objects, correctly detecting and identifying faults.
Tesla is based on cameras, and Vision-based relies on camera data. Therefore, these algorithms parse pixel-based video to detect vehicles, pedestrians, and other obstacles in the environment. Algorithms can be detected using geometry, optical flow, color, or other image features.
Waymo mentioned in the 2017 technology blog that the key to lidar is that：
Our self-developed laser radar distinguishes between real pedestrian and humanoid posters, which recognize three-dimensional shapes, detect stationary objects, and accurately measure distances.
So the core problem of the Tesla camera-based perception is that it gives the camera and lidar a similar ability to recognize 3D space objects. In other words, the ability of the automated driving system to recognize and understand image depth information is provided by cutting of different pixels.
After Elon fired a day at the laser radar. In a study by Cornell University called “Pseudo-LiDAR from Visual Depth Estimation, which narrows the gap in autonomous 3D object detection,” researchers used a low-cost, low-resolution stereo camera to pass the budget. The depth pixel back projection 3D point cloud achieves a significant increase in point cloud quality, with an accuracy from 10% to 37.9%, quickly approaching the average accuracy of the laser radar by 66%.
Considering that the research is based on a 400,000-pixel camera, there is a huge gap between the most advanced and highest-resolution camera systems, and there is no application of model distillation or anytime prediction to improve detection accuracy and speed. The conclusion is that the improvement of this research is unprecedented and has broad prospects and potential.
So how did Tesla do it?
On the day of Tesla’s autopilot investor, Tesla’s official Twitter released such a video, “With this 3D reconstruction, the Tesla vehicle can be shot in just a few seconds from 8 cameras. Collect a lot of depth information.”
This is similar to Cornell’s visual perception research path.
Of course, there are still a wide range of problems with autonomous driving that require scientists and engineers to overcome. But more often, this is a common challenge for lidars and cameras.
For example, under the condition that snow is completely covered, the camera’s ability to recognize lane lines, road boundaries, road signs, and obstacles will be greatly reduced, but the density of snow will also affect the reflection effect of the laser radar, resulting in Phantom obstacles. Thereby interfering with the perceptual capabilities of the lidar.
Therefore, whether it is based on multi-sensor fusion of laser radar or camera-based vision, it essentially gives the machine the ability to perceive the environment with human beings. That’s why the difference between Waymo and Tesla is not of much value, and why the perception of autopilot different technical routes will ultimately be the same.
There are still some things to watch, Waymo and Tesla, who wins first?