TU-TESL

• 22

At this stage, the tide is sweeping the world, and the major automobile enterprises are actively developing the auto driving function of the automobile, but their respective directions are not the same. Different from, waymo and other automatic driving companies, they adhere to the pure vision route, that is, using 8 cameras with lower cost for identification. Low cost hardware plays a crucial role in helping Tesla get a lot of data. Tesla's system written in pytorch language can automatically process data and upgrade itself. Tesla's autopilot is able to iterate quickly because of the perfect combination of hardware and software. Andrej Karpathy, director of artificial intelligence and computer vision at Tesla, made a speech for about 10 minutes, introducing Tesla how to train its autopilot system to achieve self evolution. Unlike other companies, Tesla's autopilot scheme is not based on lidar or high-precision maps, but uses a computer vision system composed of eight cameras to identify the road environment with the images taken by the cameras. At the same time, all models of Tesla are equipped with automatic driving hardware, and gradually open its automatic driving function to car owners through the form of software update, so that Tesla can obtain a large number of real data. Up to now, Tesla's navigation on autopilot has traveled 1.6 billion kilometers and automatically changed lanes 200000 times in 50 countries. The intelligent calling function has also been pushed to many regions around the world, and the car owners have actually completed more than 500000 vehicle calls. If these data are processed manually, it is hard to imagine how long it will take. Andrej karpath introduced the concepts of pytorch and torch here. It is understood that torch is an open source machine learning framework based on BSD license, which is mainly used for in-depth learning of images and videos. Pytorch is a torch machine learning framework written in Python. Compared with Torch, PyTorch has higher compilation and operation efficiency, and has laid a good foundation for Tesla's autopilot system. Based on pytorch, karpathy's team developed a computer self-training system called "vacation planning" in-house. Tesla's autopilot system can not only collect data, but also process data and achieve self evolution. This has a strong demand for the data processing logic and processing speed of the system. Karpathy showed us how the system works. Firstly, the system will decompose the collected images into several independent objects, such as traffic lights, curbs, road signs, moving objects, etc. All these data can be combined to describe the information of the current road completely. Next, the data of each object has a corresponding processor for analysis and processing. The data processing of these processors is collaborative to avoid data isolation. At the micro level, the system can monitor the data of an object; at the macro level, the system can summarize the data of each object and restore the road conditions at that time. Karpathy also introduced a feature of the system: the eight cameras of the car also work together. In many cases (such as distance and depth judgment), a single image cannot provide enough data. Multi camera collaboration will help the automatic driving system to understand the scene better. In the process of driving, for the same scene, 8 cameras will take a total of 4096 photos with different distances and focal lengths. These photos will eventually be processed in the above way. Tesla's training system eventually formed a multi camera, multi-sensor system. This system can not only analyze the road environment outside the vehicle, but also collect the treatment and results of the vehicle in response to various situations. First, the low-cost hardware is used to complete the popularization, then the delivered hardware is used to collect data, and finally the reasonable programming language is used to train the machine to independently process data and complete the self upgrading. Tesla's every move was brilliant. Large sample data and strong data processing capabilities are the two core strengths of Tesla's automatic driving system. Based on these two points, I am more optimistic about Tesla's future automatic driving system, but I am skeptical of Musk's promise to achieve full automatic driving in 2020.