Embed­ded World 2019 Flash­backs

If you want to keep in touch with the newest trends in embedded systems, Nuremberg Embedded World is the right place for you. In order to heighten your expectations, we want to share our impressions on the latest event.

More and More Chips

This year the embedded market has undergone a serious segmentation of chip manufacturers. One of the main event threads was producing chips for the deep learning and their implementation.

For instance, Nvidia’s video chipsets and their end devices for video cameras. Qualcomm in cooperation with NXP, which is famous for the achievements in vehicle automation, also follow this trend.

Companies that specialize in producing power efficient chipsets also participated in the event. A large variety of solutions with up-to-date FPGA and DSP onboard were introduced as an alternative to the standard CPU. For example, one company, produces chips with long life period for voice recognition that are as small as a pencil point, respond to certain activation words and word phrases, and has extensive battery life.

Foggy Calculations

Teradek attracted a lot of attention with the solutions based on peripheral calculations (also known as fog computing) that make the system more stable and independent from network environment and its quality. The main principle of this approach is making all the calculations on edge devices, excluding any cloud interference, that accelerates the process as there is no delay in receiving and processing the information. This approach reduces price and improves the data transfer security. Developers are now presented with a wider range of opportunities. For instance, such type of solutions can be used for elderly care without the need to install IP cameras that send the video stream to the external end user, as all the necessary information is processed and distributed to the assigned person right here and now.

Another company that uses such type of approach is Elite Vision that produces cameras with high accuracy for manufacturing facilities. They introduced camera with in-built computation module that processes the signal and provides the user with an end result of its computation. The device can be used for counting objects, scanning QR-codes, analyzing road traffic information or in any classification processes.

Such events encourage companies working in the same sphere to start new cooperation and partnerships. In that spirit, DSR is looking for camera and server station manufacturers for joint solution development.

Weapon Detection System in Public

DSR demoed a system for detecting any weapon or another predefined object (for example, products in a store) in the hands of people in a crowd. This system recognizes people and focuses on their hands. That is its peculiar feature. This product can be used for searching for people with a weapon in their hands in a small crowd. The small, power efficient chipset guarantees a long battery life period of the device and allows it to transfer the computed data and video via different types of wireless networks. This implementation of the fogging is unique because it removes the necessity of transferring data using expensive traffic. The algorithm consists of 2 neural networks: first one is responsible for identifying the wrist; the second is in charge of recognizing the object it is holding. We created a new set of data for machine learning, taught the system to accurately analyze the data, optimized the architecture and chose the right framework.

What’s Inside

In the core of the solution is the goal to make devices with such analytical power more power efficient and available to a larger number of customers without dependency on expensive hardware. That’s why our choice for a cheap machine learning device is dragonboard 410 powered by Quad-core ARM® Cortex® A53 CPU and Qualcomm Adreno 306 GPU with WiFi, Bluetooth and 3G/4G modules. Due to differentiated Quadcore CPU and GPU the development board is capable of processing several parallel tasks locally.

In addition, the system contains an Ethernet video camera, a switch, and a display. We also utilized a high performance VGA adapter produced by Nvidia and AMD for deep machine learning. Although this hardware set up meets the system requirements, its cost is still not optimal. Similar solutions are rumored to use hardware that can heat a small apartment, especially once the price of cryptocurrency falls.

This development is deep in the middle of Computer Vision – a self-learning system that is capable of learning on examples of types of weapons and human movement patterns. We are working on evolving the solution to utilize it in the area of video surveillance and public safety.