Autonomous Driving: Visual Route vs. Lidar

Recently, Tesla demonstrated how its humanoid robot, supercomputer Dojo, and pure visual route work. All reveal their technical confidence. When many manufacturers choose the lidar solution as the visual perception route, Tesla still adheres to the pure visual route.

We know that the underlying principle of autonomous driving is a combination of perception, decision-making, and execution. The perception layer uses visual sensors to obtain information about the surrounding road conditions, processes the data through the device side of the car body and the cloud, and obtains execution commands so that the car has the ability to drive automatically. Of the three basic steps, perception is the first step, which plays a prerequisite for subsequent decision-making and execution. At the perception level, there are currently two technical routes in the autonomous driving market: visual perception and lidar perception.

The lidar school believes that the camera’s dominant visual perception accuracy is not enough. If autonomous driving is to be developed above the L3 level, lidar should be used. The visual perception school believes that the environmental information perceived by the camera is rich in data, and objects can be classified and subsequently easily labeled. The most important thing is low cost, which is not possible with lidar.

Whether it is analyzed from the perspective of technology or cost, the core difference between the two solutions is whether it needs the assistance of lidar to achieve high-level autonomous driving. So, who can win the two technical routes?

Lidar vs visual, performance competition

Lidar

The lidar sensing technology is dominated by lidar, with millimeter-wave radar, ultrasonic sensors, and cameras as auxiliary. The working principle of Lidar perceiving the environment is to use Lidar to emit laser beams to measure the time difference and phase difference of the laser during the launch and retraction process to determine the relative distance between the car and the object, and realize the real-time perception of the environment and obstacle avoidance functions.

The lidar has a long detection range, high accuracy, and strong anti-interference ability. It can actively detect the surrounding multi-object environment and obtain the point cloud of the surrounding environment to build a 3D environment model. Even if the night light is not good, it will not affect the detection effect. Although lidar is not afraid of dim light, it is sensitive to weather. Rain, snow, sand, and fog, etc. affect the recognition effect of lidar. The lidar fusion high-precision map solution can effectively compensate for the high environmental dependency of the visual solution and the large demand for computing power. Its performance advantages have led most car manufacturers to list lidar as an indispensable sensing device for L3 and above autonomous driving.

Visual perception

Visual perception is a camera-led solution, and the cost of a camera has a huge advantage over lidar. The price of the camera is dozens of dollars. The lidar is several hundred dollars, which is several times that. And the camera technology is gradually mature. High-resolution, high-frame-rate imaging technology makes the perceived environmental information more abundant. However, the camera has limited perception in a dark environment, and its accuracy and safety have declined.

Compared with lidar, the weakness of visual perception is more obvious: the camera depends on the light conditions, the perception method is low in accuracy, and the dependence and requirements on algorithms and computing power are extremely high, and the barriers to data acquisition and algorithm iteration are high. In terms of performance, Lidar clearly wins. Tesla spends huge costs on computing power and algorithms, and the investment is not small.

Tesla’s logic on pure visual route

In Musk’s view, “pure visual perception is the path to real-world AI.” This is also the bottom line of thinking he pursues in solving the problem-the first principle, that is, returning to the most basic conditions of things, splitting them into various elements for structural analysis, so as to find the best path to achieve the goal.

In the process of driving a vehicle, we collect road condition information through our eyes, supplemented by brain processing. It stands to reason that autonomous driving can also drive safely through visual perception supplemented by algorithm processing. What Tesla wants to do is to imitate the ability of human vision to obtain information to achieve autonomous driving. Since the perception method of the visual camera is low in accuracy, then rely on Tesla’s unique data advantages and the ability to build computing power and algorithms to smooth out this defect.

Tesla’s advantage

While other autopilot manufacturers are still collecting data during the road test phase, Tesla has benefited from the sale of millions of cars with cameras around the world and has accumulated massive amounts of data on real road conditions. The data used for deep learning model training has long-established barriers for Tesla’s algorithm, and the accumulation speed of these data samples and the efficiency of the algorithm cannot be replicated by other manufacturers.

Tesla’s newly built supercomputer, Dojo, has powerful computing power. This supercomputer is set up for Tesla’s autopilot system. It is used to concentrate on training the entire autopilot system including Autopilot.

Tesla also made technological innovations in its cameras. Use “pseudo lidar” technology instead of lidar. Similar to the point cloud function of lidar, it can estimate the depth of the pixels in the camera to form a 3D target detection. This improved the accuracy of depth estimation, and the gap between lidar and camera began to narrow.

People rely on vision when driving vehicles. Our neural network can process distance, speed, and other signals in visual information. And Tesla’s neural network seems to be able to gradually do it. This made it impossible for latecomers to follow and copy, which also established a strong barrier for Tesla. The pure vision solution is supported by massive sample data training and advanced image processing algorithms.

Can Lidar win the competition?

The competition between the two genres depends on whether the mass production of lidar and the iteration of visual route technology are faster.

We found through data that there are more and more newly registered radar companies. The growth trend of the supply side comes from the huge demand on the demand side. Most companies engaged in L3 and L4 autonomous driving have adopted lidar.

The lidar solution, because of the safety advantages brought by the high-precision hardware performance, can be accepted by the market because of its temporary high cost. Most players accept the Lidar solution, which makes the demand side large, and its production capacity also expands. Large-scale mass production is on the way, and future costs will be further reduced due to the advantages of scale, establishing a virtuous circle.

After years of development, lidar has proven to be an indispensable sensor for high-level autonomous driving. The camera of the pure visual route is cheap but the security is worrying. It is linked to the algorithm and computing power. Tesla relies on its own massive data and supercomputers, which no one can imitate. This also means that other companies in the market cannot follow the route of visual perception regardless of the outcome.

In the long run, the visual perception route will still be controversial due to cost and safety. At present, the speed of the scale development of lidar and the speed of Tesla’s pure vision technology is unclear. But now, compared with the development of unknown visual perception technology, the lidar solution is already on the way to mass production. The development of its beautiful posture gives this faction the confidence to welcome the future with a smile.