To achieve automatic driving, first get big data

With GM, Daimler, BMW and Audi, and other mainstream car companies have announced that they will launch Level 5 autopilots around 2020. Competition in this field has entered a stage of intense heat. Internationally renowned investment bank Goldman predicts that autopilot The emergence of taxis will increase the size of the global shared car market from the current $5 billion to $285 billion in 2030.

There is no doubt that this is a big cake, so car manufacturers are busy with road tests, but the key to their ambitious autonomous driving goals is the power of analytics and artificial intelligence (AI). , establish a reaction mode of the autopilot system on the actual road, and use real-world simulation technology to accelerate the development process. This means that data engineering, management, storage and analysis become more important than ever. So, the car manufacturer should How to do it?

First, be prepared to meet the 'baptism' of massive data. Autopilot cars generate a lot of data when testing, each car will produce 6~8GB of data per second. In 2017 alone, the field created about 250 EB big data (1 EB = 1024 PB, 1 PB = 1024 TB). Automakers need advanced concepts to process this data and derive value from it.

Second, building a bridge between automotive R&D and computer and data science. Engineering is the strength of automakers, but they may not be familiar with data science. The integration of these disciplines can help automakers open new doors and accelerate their Innovation and R&D. Although the R&D department of the car company also has a dedicated data engineering team, they often need the help of experts in the fields of data science and artificial intelligence to achieve the best R&D results.

Again, efficiently process and analyze data. When autopilot cars are tested, LiDAR, panoramic cameras and radar components produce a large amount of specialized data in ADTF, ROSbag and MDF4 formats. Now there is already a lot of data available for these massive data. PB is a tool for quick access to units. In the past, it usually took several days to extract and analyze the data, but now it only takes a few minutes or a few seconds to get the results.

Fourth, selectively and targeted large data screening. Engineers can use AI technology to determine which data is valuable and which can be rejected. In general, when shooting a test scene of an autonomous car, every second will Produced 30 frames of video, but most of these videos are scenes of cars driving smoothly on open roads, and nothing special happens. Such video is not too long for automotive engineers. Autopilot The data generated by a car when it is turning, colliding, or interacting with other objects is more valuable.

Finally, optimizing the automatic generation of data. When an autopilot system makes a decision different from the human driver, it must be recorded. Similarly, when a semi-autonomous vehicle makes a decision error and is corrected by a human driver, it should also be taken seriously. This allows the engineer to optimize the system. If the autonomous driving system can be continuously optimized and improved, then the general public will be more reassured about self-driving cars.