How does Tesla's self-driving brain system learn? It will first use the camera on Tesla to build a model for the surrounding traffic participants, and at the same time add the photo data to the training set of the neural network so that humans can distinguish what it is. In this way, after forming a certain proportion, the brain? I will be able to judge for myself, such as what car, what bike, who, how far they are from me, what direction they are and the speed of their movement ... This kind of learning is not limited to Tesla's own internal test, but also includes the data accumulation of a large number of car owners in the actual driving scene. However, just like human eyes will make wrong judgments, visual algorithms also have obvious shortcomings, that is, it is difficult to infer accurate 3D real scenes from 2D plane images. For example, Tesla Motors once identified a large white truck as a cloud and ran directly into it. Take the two-dimensional human projection as a real person and take the initiative to brake; Use the stop sign on the billboard as the road sign of the car and brake actively.
Theoretically, the combination of vision technology and lidar is a perfect solution, because the image sensor in the vision scheme can obtain complex environmental information with high frame rate and high resolution, and the price is cheap, while lidar is a sensor that obtains the depth information of the target by emitting pulsed laser and detecting the scattered light characteristics of the target. It has the characteristics of high precision, large measuring range and strong anti-interference ability. Companies in the industry also generally believe that fully automatic driving technology should include sensors such as lidar, radar, camera, ultrasonic and thermal imaging.
Lidar can usually be divided into mechanical lidar, hybrid solid-state radar and pure solid-state lidar. The technology of mechanical lidar developed earlier and is more mature, but the driving environment of mechanical rotating parts is unstable, so it is very difficult to mass-produce according to vehicle regulations and the price is very high. The self-driving cars we see with high gyroscopes are generally mechanical lidar. The price of mechanical lidar that can realize L4 driving can be easily calculated in tens of thousands of dollars.