The visual system replaces the laser navigation ideal and full realistic bone sense

Source：Updated：2017-09-21 09:02:43

Nowadays, when it comes to self-driving car environmental sensing technology, many people think of lidar first. Indeed, compared with on-board sensors such as cameras and millimeter-wave radars, lidar has the advantages of high precision and high resolution, and has been widely used in many self-driving test vehicles. 
But cannot ignore this technology also has its shortcomings, the high cost, such as Ibeo LUX 4 line laser radar, priced at $15000, and although Google claims its independent development at the beginning of this year's laser radar can reduce the cost of 90%, still need to be $7500 per set. Such high prices are clearly unrealistic for self-driving cars that will eventually be commercialised. 
 
Therefore, how to find a lower cost environment awareness solution has become a concern for many enterprises after the existing lidar solutions. In response to this problem, at the 4th APEC conference on vehicle networking seminar, the Chinese academy of engineering has given its own answer - visual navigation. 
 
In his view, no matter how to optimize the laser radar cost, to bring the price down to such as $1000 or less - after all, the price is reasonable for production cars, are facing great challenges. In comparison, the visual navigation technology of "camera + software" is easier to achieve. 
 
Image. The PNG 
Laser radar price high visual navigation for autonomous driving 
 
Through visual navigation, as the name suggests that visual camera to capture image information, to obtain the location of moving object in space, direction and other environmental information, and information obtained by a certain algorithm processing, environment model is set up, and find an optimal or approximate optimal collision free path to realize the secure mobile, to arrive at the destination. 
 
In the technology scheme, there are two big key points - vision camera and artificial intelligence algorithm, in which the former is mainly used for access to environmental information, the latter is used to analyze data, extract characteristic information, so as to provide decision-making basis for further action. Compared with laser radar, dominated by the visual technology of environment perception solutions, technology is more mature, threshold and lower development costs, so in the past two years with the maturing of computer vision technology, and the Internet, artificial intelligence, the rapid development of emerging technologies such as cloud computing, get the attention of more and more automated driving related enterprises, is one of the most representative enterprise tesla. 
 
As tesla's CEO, musk has said publicly that tesla will not use lidar, which is too expensive. After the tesla Model S, on the Model, Model 3 X, really did not see the figure of laser radar, these car installed only a certain number of cameras, millimeter wave radar and ultrasonic sensors, to drive the Autopilot Autopilot. Even in 2016 tesla motors with cameras sensors failed to correctly identify the driving environment caused a traffic accident, after the controversial, the company still is not the meaning of "let go" of laser radar, this from tesla repeatedly this year to upgrade its driving assistance systems can be seen. 
 
Image. The PNG 
And tesla, like tesla, relies on computer vision technology for self-driving businesses and Tucson's future. The self-driving startup, founded in 2015, is also at the core of its self-driving solutions, with low-cost computer vision and supplemented by artificial intelligence algorithms. 
 
"Because we need to think about how to sell things, we're going to start with a relatively low price." This is the future CTO hou of Tucson. In his view, lidar has not yet been mass-produced and expensive, and cannot generate enough value at the moment, but Tucson does not rule out using lidar after cutting prices. 
 
But, as Dr Gawain says, the cost of lidar is so difficult that it is impossible to know when prices will fall to meet the demand for mass production and will be accepted by most companies. As the autonomous areas, on the other hand, companies to promote its product production process, is not much time for these companies, all on a laser radar technology "into", to look for other, more feasible, in the short term may fall to the ground. 
 
Although visual navigation still has many problems to solve 
 
Visual perception as a low cost and with the aid of big data will be able to solve the problem of technical route, although compared with laser radar, has many advantages, easier to drive self-driving cars commercial. But the route itself faces some technical difficulties. 
 
"The algorithm of AI decision is key to the visual perception of the navigation problem." Academician gao wen said. "In the case of visual distance, when the visual camera input some environmental image, we can reverse the distance of the car from the house, the pedestrian, the signal lamp by relevant calculation. Therefore, the algorithm must be good enough to measure accurately and accurately. 
 
Image. The PNG 
However, in the actual working conditions, the use of visual ranging often shows the phenomenon of position drift. Because the visual range relies on the camera to obtain different environment images, and then compared with the original map, and calculated the distance through related algorithms. In the process, if the data processing speed can't keep up, or the algorithm is not good enough, there will be a drift and a certain gap with the actual result. It will need to have special algorithm to solve the problems of the drift, such as by feature matching, extract some data about the "features", to compare the difference, so as to detect whether there is a drift, drift and position, and then reverse to correct. In addition, it can also help the vehicle with precise positioning through the idea of global optimization. 
 
Another problem is that visual navigation requires a higher level of illumination, unlike lidar, which does not require light, and can detect the distance from the car. Visual navigation due to depend on camera to collect environmental information, and the camera itself does not shine, so the light is bad, need to use auxiliary light to light, like the human eye in the middle of the night also need to turn on the light to see the surrounding environment. 
 
From this point of view, the visual navigation system of the future must also solve the navigation problem under the poor light condition. In addition, there are also weather, congestion and various emergencies, which will also affect the normal performance of visual navigation. Last may happen tesla hit the big white van is the best example, according to the tesla explanation for the accident, two cars crashed into each other when tesla is backlight driving, the strong light of tesla carrying cameras caused dry scratching, and at the time of the encounter strong light trucks of the white body, can't recognize by camera, caused the accident. It can be seen that visual navigation sounds beautiful, but it is also a long way to go.