"Internet auto". Intel products and tools

We continue to talk about digital and computer devices in cars. Today we will talk about unmanned cars as a global goal, towards which the whole community of technological industries is moving. There is a lot written on this topic, and on GeekTimes including, in this post we will look at the problem from the point of view of the “Internet of things”, and at the same time we will present two new Intel software products that will help those who overcome the hard way to create a car without a driver.

More data, good and different

As mentioned in the previous article , unmanned cars can be considered as a generator of the most powerful information flows. The speed of these streams was also given - about 2 TB of data per hour. There are more impressive estimates - up to one petabyte per hour. This includes sensor data, traffic and environmental information and so on. Obviously, it is impossible to exchange such data volume “over the air” in real time, and even to store it locally is problematic. The only way out is to keep only the most valuable, as now the existing models do. What can be done with the received information?


Simplified unmanned car architecture
')
A very promising application of this array of telematics suggests itself: use it to train an artificial neural network, so that in the future it will give instructions to your and other cars: how to recognize a pedestrian, a road sign, etc. Thus, the drone will be able to "resolve" situations in which he personally did not previously fall. In addition, the same array is the most valuable source material for various kinds of statistical calculations, which can show, say, the averaged autopilot driving style, the possibilities for increasing productivity and economy, and much more — the big date is capable of much. Well, it will actually happen both that and another, together carrying out the process of self-improvement of the artificial brain driving your car.

It sounds beautiful and seductive, but in reality there will be a myriad of problems on this path, and we don’t even guess about most of them. Although the list of preliminary questions can also be impressive. Take at least the organizational part. Will it be possible to overcome separatism of vendors and create a single “big date” for all UAVs? Who will she manage in this case? Or, here's how to ensure that the artificial mind firmly kicks its feet on the real world, without flying away to the empyrean? After all, we trust him with our lives. There are many questions, but let's see what can be done today.

Little future now

Wind River Helix CarSync Performance Diagram

The first news came from where they waited. Wind River Company (part of Intel), on the software of which the Opportunity rover and Juno research station, naturally, could not ignore the topic of automotive automation. Moreover, it has been developing it for a long time, the result was the Wind River Helix product family, which includes cloud services, auto monitoring and control systems, and autonomy elements. Now this list has been supplemented with a new component of Wind River Helix CarSync , which is a framework for remote control of the vehicle's hardware and software filling, as well as the data stored in it. That is, we got a layer ready for use between the data produced by the machine and various kinds of miracle services we dreamed of above.


Intel Deep Learning Training Tool Interface

Intel itself presents an equally interesting new free product - Intel Deep Learning SDK , a set of tools for developing, training and deploying software related to a deep study of data. The kit includes: Intel Deep Learning Training Tool, which provides a convenient graphical environment for installing, configuring, and implementing deep learning algorithms, as well as the Intel Deep Learning Deployment Tool, which improves the performance of learning algorithms on Intel platforms. Functions Training Tool:

• Creating a training dataset
• Setting up a model with automatically optimized hyper-parameters
• Launch and monitor various candidate models
• Visualization of performance and accuracy of training

Deployment Tool features:

• Import of trained models from all popular deep learning frameworks, regardless of the training hardware platform
• Compress models to improve performance, reduce space and increase transmission speed
• Generation of estimator code dependent on the hardware platform (C / C ++, OpenVXgraphs, OpenCL *, etc.)
• Seamless integration with system software stack and applications

In addition to these tools, Intel developed Intel MKL-DNN — high-performance open-source APIs for deep learning software in conjunction with the Intel MKL library. MKL-DNN is available now, the release of the Intel Deep Learning SDK is scheduled for October, here you can sign up for a preview.

Probably, it is not worthwhile to list all Intel hardware and software that can be successfully used to optimize, increase productivity, high-load computing — in general, everything that is needed in the process of creating and operating unmanned cars — let's manage this picture here.



There is no doubt that over time, autonomous vehicles will become the main means of our movement. But if by that time we will have normal roads, science cannot assert with certainty. However, this is offtopic.