In the mid-2017, during Shenzhen municipal CPPCC, Frank Wang, the CEO of DJI, aired his opinion of unmanned aerial vehicle (UAV). Speaking of UAV regulation, most measures are bans, which greatly limit the use of UAV as well as the application of new technologies.
With the increasingly rigid air policy, DJI is going to explore land transport.
But it has no intention to manufacture cars, instead, it explores self-driving.
According to the news, DJI is currently recruiting auto test editors. During the communication with DJI HR, he said: You have no idea how big business DJI has. From a large scope, all is about AI. …… As these are special projects, we haven’t recruited openly. …… It is said that nowadays, we could get auto resources of four to five brands in a week.
Signs show that DJI is doing business R&D related to smart travel. DJI is the absolute leader of UAV. Its access to automatic driving is definitely a big news for the artificial intelligence and automatic driving boom. Technically, visual recognition technology accumulated by DJI in UAV also will become its core advantage in the field of automatic driving.
DJI’s ambitions for “driving” dates back to two years ago
In 2015, the DJI R&D center in Silicon Valley hired Darren Liccardo, who was responsible for the automatic driving technology in Tesla, and Rob Schlub, Apple’s senior engineer (who is mainly responsible for the antenna design). Liccardo will serve as global engineering VP of DJI and Schlub will serve as global R&D VP.
Darren Liccardo is a talent in both automatic driving and unmanned aerial vehicles. Before joining DJI, he was the head of Tesla’s automatic driving team, and the head of BMW’s self-driving R&D team. Earlier, when Liccardo was working in Crossbow Technology, a smart sensor enterprise, he developed inertial navigation system for aircraft which was the first of its kind that certified by FAA.
Though Tesla’s automatic driving in essence is a high-level driving assistance, Tesla realizes this function mainly relies on perception, or the visual recognition of cameras. It is also one of Tesla’s leading parts in automatic driving.
DJI is a UAV company, while in essence, DJI has abundant technology accumulation in terms of visual recognition. And automatic driving is basically transferring application scenarios.
DJI is actually an AI company with visual recognition at the core
Automatic driving can basically be divided into three main parts, namely, perception, decision-making and control. And perception is the most basic automatic driving technique.
At present, the mainstream automatic driving autos studied by all companies mainly use three kinds of sensors to obtain information around the autos: camera, millimeter-wave radar and lidar. Each has its advantages and disadvantages. These three sensors record surrounding information and upload it to on-board computer to calculate. When on-board computer has a command of surroundings, it will give automatic driving decisions based on automatic driving algorithm and then control the drive.
In June 2015, DJI launched a new smart obstacle avoidance system called Guidance. The system includes two recognition mechanisms, namely ultrasound and machine vision, which help UAVs to perceive its surroundings, to avoid obstacles, and to position itself without GPS.
Functionally, is it somewhat similar to the sensor of a self-driving car?
Automatic driving is far more difficult than air travel
Technically, the vision identification and obstacle avoidance of automatic driving is far more difficult than that of UAV.
On the one hand, DJI UAV mainly uses camera as sensor, whose biggest problem is easy to be interfered. Generally, self-driving car uses millimeter-wave radar to avoid obstacles instead of camera. In this way, the stability and reliability of the obstacle avoidance system could be ensured. The entire sensor scheme of automatic driving plays its role based on the application characteristics of different sensors.
On the other hand, environment mapping and route planning of air travel is far easier than that of road traffic. Road traffic has complex traffic rules. Thus, system could make decision only after the sensor conducts a comprehensive perception on the surroundings. Therefore, although DJI has sufficient accumulation on the visual recognition technology, it is still hard for DJI to apply visual recognition technology used by UAV to automatic driving schemes.
But if DJI is really developing automatic driving, what technical details DJI is good at or suitable for?
Let’s go back to the product that DJI released. After the release of Mavic Pro, a folding UAV, a visual engineer in DJI told media: in fact, the core of Mavic lied in computer vision.
Compared with past products, Mavic Pro has more functions such as gesture selfie, object recognition, visual tracking including parallel tracking, focus tracking, auto wraparound and accurate landing. All of these functions try to solve the core and most difficult problems in the field of computer vision and robotics. And the R&D of automatic driving also need to face these problems.
DJI’s UAV mainly uses 2D cameras to realize visual recognition. It’s more difficult than 3D cameras. Besides, Mavic Pro reportedly could optimize visual recognition through in-depth learning. While in-depth learning has high requirements for computer equipment, which means that DJI needs to pay great efforts in the design of the neural network, skill training, compact and compression of models and deep optimization.
One practical problems facing the perception of automatic driving is how to deal with more data through less computing capability. DJI has achieved perception, planning and decision-making in the air. As for automatic driving, it doesn’t need to have all functions realized by visual recognition of cameras. It could directly exert the most advantageous function of camera and have millimeter wave radar or laser radar to play other functions.
From UAV to robots and from robots to AI
Whether DJI will publish automatic driving scheme or not couldn’t be confirmed from current recruitment information. But DJI’s layout of the computer vision R&D in recent years has a considerable overlap with the technical requirements of automatic driving.
Robot competition, SDK developer contest co-hosted with Ford Motor and the sponsor of CVPR, which is the top meeting in the field of visual recognition show that DJI obviously is not satisfied with applying vision technology only in UAV. Because this intelligent hardware doesn’t have a wide range of application scenario. If existing three AI directions have clear applications, it should be: in-depth learning is the basis of large data processing, semantic understanding is for smart sound box and machine vision is the essential technology of automatic driving, and also the important impetus that may change future traffic.
This article originally appeared in Huxiu and was translated by Pandaily.
Click here to read the original Chinese article.