AI and Embedded Vision Trends of Embedded World 2020

AI and Embedded Vision Trends of Embedded World 2020
Embedded World 2020 was affected by the cancellation of numerous companies due to the coronavirus. Nevertheless, there were numerous interesting new products for embedded vision and AI to be seen at the fair, which are briefly presented in this follow-up report.
Bild 1: Auf der Embedded World 2020 fand die Podiumsdiskussion 'Embedded Vision Everywhere' statt. Anwender (Stil, CST) und Hersteller (Allied Vision, Congatec, Cubemos, Vision Components) diskutierten dabei über Trends und Entwicklungen sowie die Rolle v
At the Embedded World 2020 the panel discussion ‘Embedded Vision everywhere’ took place. Users (Stil, CST) and manufacturers (Allied Vision, Congatec, Cubemos, Vision Components) discussed trends and developments and the role of AI. (Image: TeDo Verlag GmbH)

With reference to the spread of the coronavirus, around 200 companies had withdrawn their participation in the trade fair at short notice, so that in the end there were over 900 exhibitors. The number of visitors was therefore significantly lower than in previous years, at around 13,800. The embedded awards were presented on the first day of the fair. The winner in the embedded vision category is DC Vision Systems (www.dc-visionsystems.com) for the DC-SVP (Stereo Vision Processor). This combines a stereo camera and electronic control unit with an integrated processor and a hardware accelerator based on FPGA technology. A stereo image processing pipeline is implemented in the firmware. This enables real-time measurement of 3D geometries directly on the camera. In addition, Enclustra (www.enclustra.com) and hema electronic (www.hema.de) announced a cooperation agreement at the fair. The Swiss FPGA specialist will offer its range of standardized FPGA modules via a branch office at hema’s location in Aalen from May 1.

Camera(-modules)

The system architecture of the intelligent camera series iam from NET (www.net-gmbh.com) with CPU and FPGA on one chip enables higher system performance. With the additional FPGA resources, neural networks and conventional algorithms can be used more efficiently for image processing and users without VHDL knowledge can use the FPGA resources for their own solutions for the first time. With the MIPI and FPD-Link cameras from The Imaging Source (www.theimagingsource.com), in combination with the Nvidia Jetson Nano SDK, an entry into embedded vision and AI projects is possible. With pre-trained neural networks from Nvidia and MVTec, object recognition tasks can be realized. With an FPD-Link III bridge, consisting of a serializer board on the camera side and a deserializer board on the Nvidia Nano side, cable lengths up to 15m are possible. The Quad-Core processor Snapdragon 410 from Qualcomm, which is integrated in a board camera from Vision Components (www.vision-components.com), is clocked at 1.2GHz. The camera offers a GigE interface, twelve GPIOs, 1GB RAM and 16GB flash memory. An SD card slot and various other interfaces can be added via additional boards. The Alvium MIPI CSI-2 camera 1800 C-040 from Allied Vision (www.alliedvision.com) is equipped with the Sony IMX287 sensor and offers a resolution of 0.4MP and up to 292fps. Equipped with the Sony IMX273 sensor, the Alvium 1800 C-158 offers 1.6MP and up to 154fps. The new phyCAM-M camera modules VM-016 and VM-017 with MIPI CSI-2 interface from Phytec (www.phytec.de) achieve up to 60fps at full 5MP resolution. Basler (www.baslerweb.com) has cancelled its trade show appearance at short notice, but has already introduced a camera module that is compatible with NXP’s i.MX 8M Plus processor series. The latter has an ISP that provides dual camera support and enables real-time processing for pin-sharp images. A neural network accelerator delivers up to 2.3Tops. For the Framos Ecosystem sensor module (www.framos.com), numerous sensor modules (FSM) have been added for various applications. For all new FSM sensor modules, corresponding sensor and processor adapters (FSA, FPA), drivers and reference circuit diagrams are available. Spectacle-free 3D digital microscopy is made possible by the Aegolus System, which consists of a 3D camera system with two cameras and a 3D monitor that can be used without 3D glasses. The camera heads from Solectrix (www.solectrix.de) each deliver up to 30fps in 4K UHD resolution. The processing time from image acquisition to display is below 50ms.

Bild 2: Das Dual-GPU-AI-Computing-System GPC-1000 von Vecow ist mit zwei GPUs ausgestattet und bietet eine Leistungssteigerung von 37% gegenüber Kaby Lake-IPCs.
The GPC-1000 dual GPU-AI computing system from Vecow is equipped with two GPUs and offers a 37% performance increase over Kaby Lake IPCs. (Image: TeDo Verlag GmbH)

Artificial Intelligence

Deep learning networks are often considered a black box because users do not know what happens to the data during the analysis process. It is therefore difficult to correct errors in case of misclassifications. The Gradient CAM (Gradient Weighted Class Activation Mapping) tool from MVTec (www.mvtec.com) implemented in Halcon version 19.11 helps to analyze which parts of an image had a strong influence on the classification by means of a heatmap. Wahtari (www.wahtari.io) offer a complete solution, from data collection to cloud-based training on their own GPU server in Germany. The company generates customer-specific AI models and delivers them on tailor-made hardware with integrated AI accelerators. Less than 48 hours elapse from the initial inquiry to the ongoing proof of concept, and all this for a few hundred euros. Vecow (distribution plug-in www.plug-in.de) presented the expandable dual GPU AI computing system GPC-1000, which offers the power of two GPUs. It is equipped with Intel Xeon/Core-i processors of the 9th generation and the Intel C246 chipset, two Nvidia Tesla/Quadro/GeForce or AMD Radeon Pro/Radeon graphics cards. The GPU system has four PCIe slots. The devices support a power budget management with 1,500W. In edge computing, up to 750W is provided for each graphics card. The Vizi-AI development kit from Adlink (www.adlink.com) allows users to scale to other AI products in the Adlink portfolio. The kit includes an Intel Smarc module, a Movidius Myriad X VPU and 40 pin connectors. Users can start with the low-cost Vizi-AI and choose their processor type (CPU, GPU, VPU…) later. Neousys (www.neousys-tech.com) has introduced the Nuvo-8108GC AI GPU computing platform. This has a 250W Nvidia Turing graphics card with up to 14TFLOPS in the FP32 and an Intel Xeon E or 9/8-Gen Core processor integrated. ICP (www.icp-deutschland.de) completes its portfolio of AI accelerator cards with the Mustang-M2BM-MX2. In addition to Mini-PCIe and PCIe based solutions, an M.2 PCIe plug-in card variant is now also available.

IPCs & Boards

A virtual Giga-FPGA (Hexa-Core) is made possible by the Multi-FPGA-Board from ABS (www.abs-jena.de). The concept is based on a board with two FPGAs ECP3-150 and one LFXP2 as programmable controller device. The board can be extended by up to four FPGA plug-in modules. The plug-in modules are linked to the adjacent FPGAs via high-bit-rate SerDes units. The C-Vision Kit from Aries Embedded (www.aries-embedded.com) is a design platform for embedded vision and AI. The core is the MCV SoM on the MCVEVP baseboard, which can optionally be expanded with a 7-inch TFT display and capacitive touch screen. The C-Vision adapter allows two Basler dart LVDS cameras, a miniPCIe card with Movidius Myriad X chip, and a monitor to be connected via an HDMI slot. Congatec (www.congatec.de) is expanding its embedded vision offering with a solution platform for the NXP i.MX 8 processor series. The application-ready arm platform integrates all components required for MIPI camera support onboard. The kit is based on a modular 3.5″ carrier board, is available in various Smarc CoM configurations, and comes with a 13MP Basler BCON for MIPI camera module. The VarioFlex rackmount platforms from Pyramid Computer (www.pyramid.de) offer short installation depths of 250mm and optional front I/O connectors. Equipped with numerous slot options for installation depths starting at 300mm, the VarioFlex 2U platforms offer e.g. power consumption up to 140W TDP. The Kuber series from Portwell (www.portwell.eu) features the Intel Celeron processor N3350 with 4GB (LPDDR4 2400) memory. It offers Intel Gen9 graphics with up to 18 execution units and enables the connection of 4k displays via the display port.

Bild 3: Den embedded award in der Kategorie Embedded Vision hat DC Vision Systems für die Real-Time 3D-Stereo-Kamera DC-SVP gewonnen.
The Embedded Award in the Embedded Vision category was won by DC Vision Systems for the real-time 3D stereo camera DC-SVP. (Image: TeDo Verlag GmbH)

Conclusion

Despite the numerous cancellations by some 200 companies, the exhibitors in Nuremberg were quite satisfied with the visitor response. The visitors found numerous classic machine vision exhibitors, especially in the Embedded Vision Area, which was held for the first time, who presented their embedded competence there. The main trend at the exhibition was again AI, which is very closely linked to the subject of machine vision, so that numerous Embedded Vision exhibitors and users will again make their way to the exhibition in 2021. The Embedded World 2021 will take place from 2 to 4 March in Nuremberg, hopefully without flu-like side effects.


Highlights of the Embedded World (in alphabetical order)

Real-Time 3D Stereo Camera

Thanks to its integrated processor and FPGA hardware accelerator, the DC-SVP stereo camera allows 3D geometries to be measured directly on the camera.

www.dc-visionsystems.com

White Box AI per heat map

The Grade-CAM tool helps by means of a heatmap to analyse which parts of an image had a strong influence on the Deep Learning classification.

www.mvtec.comc

AI integration for beginners

Even users without FPGA or VHDL knowledge can integrate their own AI application solutions into the iCam smart camera thanks to a special SoC architecture.

www.net-gmbh.com

Dual GPU power

The GPC-1000 expandable dual-GPU-AI computing system offers the power of two GPUs and 1,500W power budget management.

www.plug-in.de

Individual AI in 48h

Wahtari generates custom AI models. It takes less than 48 hours from inquiry to proof of concept.

www.wahtari.io

TeDo Verlag GmbH
https://www.tedo-verlag.de/

THIS MIGHT ALSO INTEREST YOU