Yesterday I drank with two brothers, and one of the brothers also worked for a well-known domestic automotive software service company, and has been doing the automatic driving of Beiqi Extreme Fox Changan Automobile. Although I shouted at my handsome brother, but the depth of understanding of the business, the algorithm of the code, and the output of the code, I am far worse than others.
During this period, a topic was discussed, that is, the solution of autonomous driving. Because he has been doing domestic business, the basic solution is millimeter-wave radar, lidar, ultrasonic radar and multi-sensor fusion perception technology such as cameras. And I talk about Tesla's pure visual self-driving solution, just using cameras to achieve automatic driving is also very powerful. He immediately retorted to me that radars of different styles in rain, snow, fog and fog can achieve penetration for 3D modeling. So how do you do 3D modeling with just a camera? Sun Jian also asked me if the battery is dead, but you said that the Tesla Vision system uses powerful algorithms to achieve automatic driving, but you said that the specific theory of Tesla Megatron system to achieve 3D modeling is indeed my knowledge blind spot. Today I started to make up for it when I had time. After you understand it, you quickly take it out and share it with everyone.
Since the audience watching my videos are ordinary people like me, I try to say some colloquialisms and want to see a deeper interpretation. I put the link in the comment area, if you are interested, you can take a look. To get back to the point, the core idea of the four departments is to simulate the recognition of objects by human driving through machines, which is basically the same set of AI artificial intelligence. Human cubs are like saying that they don't know what a dumpling is, whether they can eat or not, and they can get this ability through a lot of training.
In this piece, add the three basic elements of artificial intelligence: data, algorithms, and computing power. I don't quite understand, I can look through my previous videos. In short, people and machines, like machines, do not know what to eat and what not to eat in the next birth, they all need to feel slowly and then form habits. It's just that this habit is taken for granted by humans, and the feeling is innate. Machines need to learn slowly.
About machine learning, focus on the algorithms of deep learning. Because the color 2D pictures collected by the camera, the extraction of picture information, the existing pictures, the extraction algorithm has been very mature, and it is over.
However, when the vehicle is driving at high speed, the surrounding vehicles and road conditions are changing rapidly, and how to identify the distance of objects through images is indeed a major weakness of pure visual automatic driving.
A term called depth estimation network is actually estimated ranging, mainly through parallax prediction to obtain images collected by binocular cameras (calculating depth). The depth can be predicted by the visual difference of the two images taken separately.
To explain again, first collect a photo of the T moment while driving, obtain a depth picture of the T moment through the depth estimation curl muscle network neural CNN, and then take the depth picture to synthesize the two pictures before and after, and then take the real original picture of the two moments for comparative training. Tesla trained like 3 billion miles to get the predictions to a fairly accurate effect. At the same time, the vehicle will also remember, intelligent memory human operation in this case, repeatedly match visual content and operation, similar to human following the car, overtaking and changing lanes after recognizing the distance of the car.
Going back to the rain and snow weather or the situation of poor vision at night, it is also collected through multiple cameras, collecting a large number of data analysis and matching, and making the final decision. It's like people in fog and snow, can't see the road conditions clearly, but if there are two people with four eyes looking together, does it greatly increase the success rate of road conditions. If it is 4 people with 8 eyes, of course, in extreme cases, several eyes are not easy to make, which is also normal. Multi-sensor fusion, just like seeing is very important, mainly by touching, touching what is what, pure visual perception.
It's like just looking at what you see, and what you see is definitely having its disadvantages, but it can't be said to be completely useless. Why develop pure vision autonomous driving? Because the cost of 8 cameras of the cheap Tesla added together, there is no high cost of a lidar.
Why is only Tesla developing pure visual self-driving? Because only Tesla has millions of customers for more than a decade, and more than 3 billion miles of feeding data around the world, can invest in such a high IQ of this Tesla Vision system, this issue.
Speaking of which, another topic on the wine table in the next issue is the domestic large aircraft C919, bye-bye.
