The biggest challenge for automatic driving is that it......

Mobile equipment camera software

leifengwang· 2017-02-04 04:35:43

" with the "TeleEye" and "The Brain", there is no doubt that self driving cars have begun to take the driving ability without obstacles.

according to Lei Feng network understanding, at present, the mainstream of the automatic driving of vehicles, pedestrians or animals to identify the accuracy rate is rising, thanks to their different equipment and equipment equipped with the system. But most people did not expect is that the lightest on the road, the most quiet, the most flexible car, the car may become the biggest challenge in the future.

background: "small"

"bicycle trouble may be automatic driving system of face detection problem," the most difficult, researchers from the University of California at Berkeley Engineer Steven Shladover said. California University

visual computing expert Nuno Vasconcelos also expressed similar views, he believes that bicycles are due to the relatively small and flexible and special structure, may make the automatic face a complex calculation problem of driving a car. "The car is like a" huge monster ", but the quality of bike is much smaller, and their shape is different -- such as different shapes, different colors, different ornaments and so on.

this also explains a problem, that is why in recent years, automatic car detection accuracy of the car has exceeded the bike . Of course, there are also some reasons for algorithm training. At present, most of the road detection algorithms for automatic driving vehicles are based on the depth of the image, while the main "learning" object is the characteristics of the car, the bike is relatively small.

algorithm turn: Subversion

saying here, have resorted to a mysterious algorithm recently issued. This mystery is more than the algorithm itself, but rather interesting research team lineup. Lei Feng learned that at the end of last year, these researchers released their algorithm called Deep3DBox algorithm.

" you can see from the above, the thesis is composed of three people together to complete, one from George • Mason University, and two others, from extremely low-key to have fired the reputation in outside, the driverless car company Zoox.

know this area should be on Zoox how many have heard, the United States coordinates of Silicon Valley start-up company is committed to the unmanned vehicle research and development. In October last year, after a sum of up to 50 million financing, Zoox has reached a value of $1 billion. Although Zoox may never "propaganda department", but its technical staff team is very strong, some of the data show that some of the staff worked at Google, apple and other companies, tesla.

back to the new algorithm "three joint research and development of Deep3DBox," explained in layman's rather, this algorithm is an original way to target detection and extraction of 3D from the 2D photo pose estimation . The test data show that Deep3DBox can identify 89% vehicles, and before the record holder (academic direction) score is 70%. The above mentioned

" America commercial mathematical software MATLB in the "boundingbox" (the bounding box) is a word made definition, namely "get the properties of region", meaning is used to measure the image area attribute function . In many image recognition based on 2D images, have used this function, such as license plate recognition, etc..

in this context, we can more intuitive understanding of image recognition, 2D boundingbox and 3D boundingbox distinction. For example, using 2D boundingbox robot to grab a car in the image, with a square circle is equal to it, but the 3D boundingbox circle is a "Cube", in the space dimension, 3D boundingbox can calculate more information.

now, Deep3DBox is trying to break a bigger challenge: that is, they rely on the algorithm, the obstacles of 2D target image around the vehicle in the three dimension, to predict the vehicle route.

George • Jana Ko&scaron computer scientist Mason of the University; ecká explained, "the detection and recognition of deep learning is usually suitable for pixel mode, but we found a way to use similar means to assess the magnitude of geometric level. "Koš ecká and a special identity, which is the Deep3DBox project investor.

bottleneck: Bike problem

Deep3DBox has been a lot of improvement than the original depth of learning, but bicycle recognition is still a problem. The

team found that 's performance was significantly reduced when the algorithm was involved in positioning and orientation of bicycles, . In the benchmark test, the target recognition rate of the algorithm is only 74%, even if this has been regarded as the best results in similar algorithms, but for the perfect operation of the car is not enough. In the

orientation test, the accuracy rate of Deep3DBox to judge the vehicle is 88%, but only about 59% for the bicycle.

Koš ECKA points out that, relatively speaking, the commercialization of automatic driving

The lastest articles of leifengwang