
File photo: Autopilot test. Photo by Tomita
Beijing (CNS) -- The Second Department of the Second Academy of Aerospace Science and Industry of China (CASIC) said Monday that the department has successfully developed an "intelligent driving assistance system based on deep learning", which can realize real-time and accurate intelligent perception of the environment by using an embedded chip several centimeters in size, and has reached the world's advanced target recognition accuracy.
Public data show that the highest accuracy of intelligent driving target recognition is 90.55%, but it takes 4 seconds to process a picture; The algorithm accuracy of the two intelligent driving assistance systems is 90.05%, but it only takes 0.03 seconds to process an image.
At present, the two teams have made breakthroughs in a series of core technologies, including multi-target detection and recognition, driveable area segmentation, lane detection, etc., and formed technical barriers in deep neural network compression, deep neural network compilation tool chain and intelligent acceleration engine. Guo Rui, head of the team, said that scene semantic segmentation is to let the computer understand what it "sees" the image represents, and deep neural network can learn to extract high-level semantic features, excluding the interference of backlight, light, shadow, defect and other factors on the target detection.
Guo rui said the team's focus has shifted to the engineering and production of "intelligent assisted driving system based on deep learning", and it is expected to complete small-scale mass production by the end of this year by jointly carrying out functional tests and mass production trials with related auto factories.
In addition, the team is developing new smart sensors that combine visible light, infrared and millimeter wave radar to cope with special situations such as complex and changeable lighting in military vehicles' driving environment and forbidden lighting at night, so as to meet the needs of military vehicles for round-the-clock and all-weather driving assistance. The advanced technologies accumulated by the team in intelligent perception, intelligent decision making and other aspects are gradually playing a role in weapon system target detection and recognition, intelligent accusation and other fields. (Wang Zhen Reporter Fu Yifei)
(Reproduced in Science and Technology Daily)