Q theory is not a slap in the face. Tesla’s test of lidar may be to abandon all radars faster

0
175

Recently, a photo of Tesla Model y equipped with lidar was circulated on social media, leading to speculation that Tesla and luminar are cooperating.
According to media reports, Tesla has signed a contract with luminar, a lidar company, to use laser sensor technology for testing and development.
Did musk, who insisted that “a fool uses lidar”, hit himself in the face so quickly?
Not necessarily.
Zhu Yulong, an electric vehicle engineer, believes that Tesla is still taking “the most determined pure vision technology route in the world”. This time, the lidar revealed by foreign media will not appear in the mass production vehicles, and the test lidar is more “likely to be a truth sensor used to establish the visual training set”.
Zhang Kangkang, Ph.D. of Tsinghua power engineering and engineering thermal physics, also believes that its real purpose is “to train visual algorithms with lidar, so that it can grow rapidly, and eventually abandon all radars completely.”.
Yang Kuan, an intelligent electric vehicle engineer, holds the same view: “Tesla is more likely to use luminar’s lidar to verify its fully automatic driving function, or to evaluate the performance of lidar compared with camera system.”
On the other hand, some people speculate that Tesla may really want to apply lidar. For example, he Rongliang, a researcher at CCID think tank, thinks that the possible reasons are: first, the technical route of lidar + high-precision map has begun to show obvious advantages. Second, the price of software and hardware, including lidar, began to decline sharply.
Despite speculation, the final truth will have to wait for Tesla to respond in person.
In this question you asked me, we invited people in the industry to discuss with you.
Tesla’s test of lidar may be to train visual algorithms
@Zhu Yulong, the nickname of Tencent News “automotive electronic design” and “three electric engineers of electric vehicles”
Personally, I think Tesla is the most determined pure vision technology route in the world. A while ago, Elon Musk even thought that millimeter wave radars need to be marginalized. This time, the lidar revealed by foreign media will not appear in mass production vehicles.
From a practical point of view, if the lidar is installed, the hardware modification of Tesla will not be able to be forward compatible. Even the computing platform needs additional design modification, which involves the layout of lidar, cable connection and data processing.
Judging from previous reports and this update, Tesla’s use of lidar may be a truth sensor used to build a visual training set.
What does this mean? It is in the process of continuous training that a reference system is needed to establish the basic ability of the automatic driving system. For Tesla’s updating vision sensor parameters and determining software algorithm, the sensor with outstanding performance is often selected in the perceptual ability test. The data collected by it will be used as the “reference answer” to evaluate the target perception.
Lidar has natural advantages in acquiring 3D contour of target, accurate ranging and real-time trajectory tracking. Lidar has been widely used in ranging calibration. The lidar / MMW radar data associated with visual data is used to accurately evaluate the target information.
@Zhang Kangkang, PhD, power engineering and engineering thermal physics, Tsinghua University
This is “Tesla uses lidar on test vehicle”, not “Tesla tests lidar for mass production vehicle”.
In short, this is using lidar to achieve more accurate detection, with more accurate results to train the visual algorithm. In other words, “use lidar to train visual algorithms so that they can grow rapidly and eventually abandon all radars.”.
So far, musk has not yet said that he wants to give up the pure vision scheme.
However, I think that with the continuous reduction of lidar cost, it is very beneficial to use lidar in the automatic driving scheme to achieve higher safety.
The official account of Yang Guang is “the reference of the automobile man”, the engineer of the intelligent electric vehicle.
Why musk doesn’t use lidar:
First, the first principle is that people can drive anywhere because they have two eyes. As long as they use enough real-world data and develop and train a set of neural networks to simulate human vision, they can realize automatic driving. Therefore, musk believes that Tesla does not need lidar.
Secondly, lidar needs to be used with high-precision map, and the construction of high-precision map is very complex, which is a very expensive and challenging resource intensive work.
Third, the lidar can’t identify colors, so it still needs to be used with the camera. The expensive price and ugly appearance are also the reasons for restricting the use of Tesla.
Just from a test car does not mean that Tesla’s technology route has completely changed. I think Tesla is more likely to use luminar’s lidar to verify its fully automatic driving function, or to evaluate the performance of lidar compared with camera system.
Musk was not slapped in the face, rational view of Tesla automatic driving test using lidar.
Lidar detection accuracy is high, and the cost will be greatly reduced
@He Rongliang, researcher of saidI think tank
The realization of automatic driving technology mainly depends on sensor equipment, mainly including high-definition camera, millimeter wave radar, ultrasonic radar and lidar. The specific advantages and disadvantages are analyzed as follows:
First, high-definition cameras, like the eyes of a car, can recognize all kinds of objects in the lens. They have the image recognition function that radar can’t complete, and the price is low. The disadvantage is that it is greatly affected by the ambient light.
Second, millimeter wave radar. Millimeter wave radar has strong anti-jamming ability and high penetration. It can penetrate fog, smoke and dust. Except for heavy rain, it is hardly affected by the weather and its cost is low. The disadvantage is that the detection range is still relatively short. China mainly relies on imports of high-frequency millimeter wave radars with strong functions at one or two hundred meters.
Thirdly, ultrasonic radar has a wide range of ultrasonic scattering angle and poor measurement accuracy, which is suitable for short distance measurement. It is mainly used for automatic parking, and its measurement ability is obviously limited in high-speed driving.

Fourth, lidar. Now, the breakthrough point of the industry is focused on lidar, and high-end models are competing to install lidar first. The detection accuracy of lidar is high, which can achieve centimeter level, and the detection distance is long, more than 200 meters. The disadvantage is that it is more vulnerable to the influence of natural light. In places with strong light, the lidar will be weakened. In addition, the price of lidar is high.
However, it is believed that with the increase of assembly models, the cost of lidar will also be greatly reduced.
Tesla started testing lidar mainly because:
First, the technical route of lidar + high-precision map has begun to show obvious advantages. At present, including Baidu, Google Waymon, Weilai and Xiaopeng, the main sensor devices are lidar, millimeter wave radar, ultrasonic radar and high-definition camera to realize the automatic driving function above L3 level.
Second, the price of software and hardware, including lidar, began to decline sharply. Three years ago, a high-level solution of automatic driving assistance system for automatic driving test vehicle, which required more than 3 million yuan in software and hardware, has now dropped to more than 1 million yuan. In the future, the decline will be relatively large. Especially in 2021, it is expected that in the second half of 2021, the state will liberalize the L3 level automatic driving policy and realize mass production of key parts, which will drive the price down significantly. From the function of automatic driving system, it mainly includes radar and visual sensing equipment, computing platform, algorithm, high-precision map, positioning technology, 5g vehicle networking technology, etc.
@Finance columnist
Not only Tesla, but also Baidu, which has been dedicated to visual algorithms, has changed.
According to Baidu, “on May 18, 2021, baidu Apollo signed a strategic cooperation agreement with Hesai technology. According to the agreement, baidu Apollo will customize the laser radar with a new architecture of Hesai technology for the fifth generation of fully driverless shared unmanned vehicles. ”
A key point of Baidu’s cooperation with lidar manufacturers is: “the customized lidar of Baidu Apollo will be used for commercial operation of unmanned driving. Compared with the same type of general lidar, its performance has been greatly improved, and the cost has been reduced by nearly 50%.” In other words, the cost of lidar is greatly reduced.
Did Tesla also find a laser radar supplier with “high quality and low price”? Therefore, the role of cost reduction must be great. Lidar’s position in Tesla will gradually rise. Nevertheless, Tesla and Baidu’s exploration of visual algorithms will continue.
Safety first, lidar is the mainstream configuration of automatic driving
@Chief analyst of D Master Ai Rui
Autopilot is divided into three layers: perception layer, decision layer and execution layer.
The battle between lidar and camera is a battle in the perception layer. At present, most car dealers use both of them. After all, people’s lives are at stake, and no one dares to be careless.
The two have their own advantages, but the price difference is quite big. Musk’s idea of realizing automatic driving, with his consistent idea of reducing the cost of industrial products, is bound to want to kick off the radar.
In order to replace the radar, the technology must be perfect, and the algorithm of the decision-making level must be absolutely OK. This is the problem of occlusion and light that plagues the artificial intelligence visual recognition industry. And there is no solution to this problem, the light problem can be improved by the ability of the camera, but the rest of the recognition and algorithm is a global problem.
From a pragmatic point of view, you don’t care what it uses, the final test result is safe.