What is the purpose of introducing a pure visual autopilot scheme?

Baidu Apollo unveiled a closed-loop solution for auto-driving pure vision urban roads, Baidu Apollo Lite, at the recently concluded CPVR (Global Top Academic Conference on Computer Vision and Pattern Recognition) in 2019.

According to Wang Liang, chairman of Apollo Technical Committee, Apollo Lite is the only L4-level visual perception solution for urban roads in China. It can support parallel processing of 10-channel cameras and 200 frames/second data. The maximum frame loss rate of single visual link can be controlled below 5, realizing full 360 degree real-time environmental perception and forward-looking. The visual distance of obstacle stability detection reaches 240 meters.

Tesla should be happy with the release of Baidu Apollo.

In the field of automatic driving, only the Tesla family is a staunch supporter of pure vision. On Autopilot Day in April, Musk also openly mocked that "all people who use lidar to drive automatically are silly X", offending almost all his peers.

Now, with Baidu's support for pure visual perception, Tesla may feel warm in her heart: I'm no longer fighting alone.

Neither the Lidar School nor the Pure Visual Perception School holds selfishness.

The so-called lidar school and pure visual perception school are two different technological orientations of automobile perception of the external environment. According to the basic principle of automatic driving, the perception layer is composed of various hardware sensors, which is used to capture vehicle location information and external environment information, and then enter the decision-making and execution.

Around the choice of automatic plus sensing hardware, pure vision and lidar are formed. For a long time, companies such as Mobileye and Tesla, which adhere to First Principle, believed that the use of cameras was enough, but most companies believed that lidar was essential.

It is worth noting that since 2018, Mobileye has also begun to "revise" its own route of perception technology, adding lidar to the test vehicle, and even, Mobileye intends to do lidar in person. At present, among the mainstream companies of automatic driving technology, only Tesla is still insisting on pure vision.

From a technical point of view, both schools have their own strengths and weaknesses.

Although the detection range and accuracy of lidar are much higher than that of camera, it can not recognize the color and track the target. At the same time, the more signals processed in high-speed mobile, the larger the number of pixels will interfere with the radio and television detectors, resulting in the decline of recognition accuracy, and the need to insert additional external adapters alone. Charging, using more cumbersome.

The advantage of pure vision scheme is that the video data obtained by camera is most similar to the real world perceived by human eyes, and is also closest to the shape of human driving. At the same time, the development trend of high resolution and high frame rate imaging technology means that the environmental information contained in the image is more abundant, but the camera relies on environmental light and is easy to accept. To the adverse environmental impact, and can not directly judge the depth of field, the algorithm, computational requirements are very high.

From a commercial point of view, both factions are actually secretive.

At present, the perception ability of lidar is stronger than that of camera, which also leads to the "shortcut" of technology companies and automobile companies in order to speed up the commercial landing of automatic driving. They rely too much on and build lidar in the scheme, while avoiding the basic problems of automatic driving such as visual recognition algorithm and chip, and use Tesla AI. In the words of senior director Andrej Karparthy, "it gives people a false sense of technological progress".

One of the reasons Tesla's pure vision solutions is that lidars are too expensive. Waymo was the first to use Velodyne's lidar on its test vehicle, with a single price of $75,000, making it difficult for users to digest the cost. Tesla, after all, is a company that sells new energy vehicles to the C end, not a company that sells auto-driving technology and solutions. From the long-term development trend of technology, pure vision scheme may not be able to complete the whole process of automatic driving, but the core hardware camera of pure vision scheme and its visual algorithm must be the core of automatic driving perception technology in the future. The low-cost pure vision solution of Tesla Station is not only for current sales, but also for future trends.

Is Baidu a side team or something else?

This time Baidu released a pure visual program, is it a simple demonstration of technical ability, or is it brainwashed by Musk, and decided to form an "alliance" with Tesla, standing in a pure visual program camp? Intelligent relativity holds that Baidu's actions can be interpreted from the following aspects.

1. Grinding Your Visual Technology

As mentioned above, no matter how the autopilot technology develops, the hardware and algorithm with the camera as the core will be the core of the future autopilot sensing technology. Baidu is one of the leading technology companies in the world in autopilot research, so it is necessary to lay out related technologies in advance.

Wang Liang, chairman of Baidu Apollo Technical Committee, stressed that Baidu's determination to devote resources to developing pure visual perception solutions does not mean abandoning the existing technology route based on lidar, but fully realizing the necessity of the unmanned system true redundancy in the process of technological practice and deciding to adopt the technology of pressure circumferential vision. Tamping multi-sensor fusion sensing framework.

To put Wang Liang's words more plainly, in the traditional perception solution based on lidar and supplemented by vision, the problems and defects of visual perception are exposed under the cover of the advantages of lidar.

Please read the Chinese version for details.