Demonstration of deep learning by the development unit "NVIDIA Drive PX 2" that realizes automatic driving car is like this

A unit that NVIDIA can recognize the appearance of the outside world using deep learning in real time and develop automatic driving car technology "NVIDIA Drive PX 2We presented at the conference held in accordance with CES 2016. If you look at the demonstration using deep learning technology, you will understand the awfulness of NVIDIA Drive PX 2 that it took months to finish work that took months.

Live: NVIDIA's 2016 Las Vegas CES Press Event
http://blogs.nvidia.com/blog/2016/01/03/ces-las-vegas-event/

Automotive Innovators Motoring to NVIDIA DRIVE | The Official NVIDIA Blog
http://blogs.nvidia.com/blog/2016/01/04/automotive-nvidia-drive-px-2/

Conference started at 11:00 on January 5, 2016 in Japan time.


Jen Seung Juan CEO appeared.


Technical competition of automatic driving car (self driving car) is occurring all over the world.


NVIDIA, which supports automated driving car development utilizing the computation processing capacity using GPU, introduced the new system "NVIDIA Drive PX 2".



Development of an automatic driving car is an innovative technology that changes the concept of a moving body. It is expected to realize a comfortable urban environment such as reducing traffic accidents that lose a lot of lives, creating new personal mobility, and eliminating shortage of parking lots.


There are two visions to the development of automatic driving technology by NVIDIA. One is a technology that allows a driver to move to a destination without operating it. And the other is a car that can move without a driver. This is likely to be useful in transportation and dispatch services.


Techniques necessary for automatic driving cars are to process maps and surrounding environment with recognition information obtained from sensors. This process is to draw a loop like the figure below continuously.


Because the amount of information to be processed is enormous, it is extremely difficult to develop an automatic driving car.


Especially, it is extremely difficult to accurately judge the circumstances that change everyday. Not only cars but also undefined elements such as pedestrians, bikes, bicycles, neglected objects, stray dogs are packed in the town. Technologies that recognize these are the keys to the realization of automatic driving cars.


More specificallyDeep learningIt is thought that technology will be solved. A neural network like the human brain processes visual information and performs machine learning.


No programming is required for this machine learning. The neural network has evolved by inputting various information and training the computer itself (deep learning). With GPU of NVIDIA Drive PX 2, training speed has been increased from 30 times to 40 times. It took only a day to complete what took months to date.


IT companies such as Google, Microsoft, IBM, Baidu and others participate in AI development competition. According to Google and Microsoft, the neural network has exceeded human ability in speech recognition and IQ test in the past year. Speeding up training speed plays a decisive role in the speed of AI development.


Developing AI using NVIDIA's GPU unit is not limited to large companies such as Google and Microsoft, but it is diverse from startup to small business.


The key to the automatic driving car is deep learning, but the GPU handling it is holding it.


NVIDIA proposed an end-to-end deep learning platform. With a single NVIDIA platform anyone can participate in the development of deep learning.


The processing capacity of NVIDIA's deep learning rises to the right. He said that he reached the world's top level in just a few months.


Here Mr. Mike who is in charge of image processing technology development appeared.


I analyzed the image of in-vehicle camera using NVIDIA Drive PX 2 and showed how to recognize a car on the road in real time.


The object learned for training of the neural network is 120 million pieces. Without GPU acceleration, it took several years to complete it.


It is not limited to cars that can be detected. Pedestrians and signs can also be identified.


It is also perfectly distinguishable from the road.


We will also test on the expressway effortlessly.


He said that he is doing it with deep learning without writing all the program code and it is much cheaper than hiring an excellent programmer.


Here is an introduction of development examples. Audi learned German road signs in less than 4 hours and it is possible to identify with 96% accuracy.


Daimler is perfectly successful in image recognition.


JapaneseRobot TaxiEven use NVIDIA platform.


BMW, Toyota, Ford etc. are also using the NVIDIA platform to develop automatic driving technology.


The flow of deep learning is input to the car-mounted NVIDIA Drive PX 2 after the processing result by GPU is raised to cloud-based NVIDIA DRIVENET. Furthermore, the learning contents by NVIDIA Drive PX 2 are processed and accumulated and used in NVIDIA DRIVENET, and it evolves continuously.


In deep learning in automatic driving car development, lidar (rider) attached to the front and back and the roof a total of 4 ... ...


We will obtain and obtain information from a total of 6 cameras mounted in front, back, left and right.


Four riders scans four times per second, real-time 3D information conversion for each rider around the rider.


Three-dimensional grasp of surroundings in real time. Of course, this work is 3D-processed according to the running speed of the car, so it seems that extremely high computing capacity is required.


Fisheye cameras are ants before and after.


By integrating all these information, real-time sensing is possible.


By tracking the front and rear of the body, you can track other cars running in front of and behind the vehicle.


Use the GPS to locate a rough driving position.


Furthermore, it is aiming to specify the position information at the centimeter level in which lane the surrounding recognition technology is.


In addition to these, if you also capture highly accurate map information ... ...


Automatic operation mode is possible.


It is necessary to realize automatic driving cars by ultra high speed processing of information of all external circumstances, and NVIDIA realizes this by ultra high speed parallel operation by GPU.


Mr. Justin appeared here.


Explain car information processing.


In the image below, the front and rear camera images are displayed up and down, and the information on the outside world of the automatic driving car is displayed on the center infometer.


When I recognize a car approaching from front and back, left and right, I am warning with a yellow display.


Automatic Drive Utilizing Deep Learning After explaining the car development, finally introducing the hardware of the new platform "NVIDIA Drive PX 2".


NVIDIA Drive PX 2 adopts 16 nm FinFET which further refined 28 nm process adopted in TITAN X. Equipped with 12 CPU cores capable of computing 8 teraflops.


The GPU adopts Pascal and boasts 150 times the computing power of the MacBook Pro with a size fitting in the car's trunk.


Show off the real machine here.


NVIDIA Drive PX 2 main board.



Water cooling type cooling mechanism is adopted as an option.


Automatic driving by introducing NVIDIA Drive PX 2 Public road driving test of a car will be held in 2017, the first thing seems to be Volvo.


The development unit "NVIDIA Drive PX 2" will be offered to development partners in the second quarter of 2016 (April to June) and general sales will be scheduled for the fourth quarter of 2016 (October to December).