Although there is no universally valid definition for the term 'embedded vision', it is still a common description: compact image processing systems based on customized camera modules are integrated directly into machines or devices, where they use tailored computer platforms and low power consumption to provide intelligent image processing in a wide range of applications without the need for a classic industrial PC.

Peter Keppler, Stemmer Imaging: "The topic of software is one of the decisive keys on the way to the optimum image processing system."

© Stemmer Imaging

A distinction must be made between different types of embedded vision systems: The easiest distinction is between embedded vision systems and classic image processing systems, as Peter Keppler, Director of Corporate Sales at Stemmer Imaging, explains: "Classic image processing systems work on the basis of industrial PCs that are freely programmable via special image processing libraries. Images are captured using cameras equipped with suitable optics. Illumination, which should be optimized for the respective application, ensures sufficient and application-specific illumination of the test objects." After image acquisition, the camera data is forwarded via suitable interface cables to image acquisition cards, which coordinate the actual image processing on the computer's CPU. With some of these frame grabbers, image pre-processing is carried out to reduce the load on the host CPU. The results of the evaluation provided by the system are usually used for quality assurance of manufactured goods.

Embedded PCs differ from these IPC systems in that the functionality of the image acquisition cards is already integrated into the embedded PC. Like classic industrial PCs, they also enable the connection of external standard cameras with all image sensors available on the market. Based on Windows Embedded operating systems, Embedded PCs are also freely programmable and allow flexible adaptation of the systems to the respective requirements via special libraries for industrial image processing. The connection to the machine is established via proprietary bus adapters or special industrial Ethernet cards. Examples of embedded PC systems include the 'CVS Image Station Compact' from Stemmer Imaging, the 'IPD GV' family from Teledyne Dalsa and the 'Matrix' series from Adlink.

Smart cameras and vision sensors

Smart cameras and vision sensors combine the camera, image capture, the processor for image evaluation and the I/O interfaces as well as lighting and optics in a single housing.

© Stemmer Imaging

Smart cameras and vision sensors go one step further: They combine the camera sensor, image acquisition, the processor for image evaluation and the I/O interfaces in a very compact, robust housing. Vision sensors usually have graphical 'point and click' parameterization. These systems often also work with integrated lighting and optics, which simplifies the application but also reduces flexibility: "Such systems are usually optimized for specific applications and do not allow a switch to a completely different application, for example from pure presence monitoring to measuring or reading tasks. Another limitation is the limited range of image sensors that can be used in such products," Keppler points out.

There is no precise conceptual distinction between smart cameras and vision sensors; typical representatives of this class are the stand-alone products of the 'InSight' family from Cognex and the 'Boa' series from Teledyne Dalsa.

Specialized in one task

Fully integrated image processing systems such as 'deep embedded vision systems' have been specially developed for a specific task and are not freely programmable.

© Stemmer Imaging

For fully integrated image processing systems that can also work without an operating system, Keppler suggests the term 'deep embedded vision system', which is not yet established on the market: "Such systems are specially developed for a specific task and are not freely programmable. The communication options of the systems are already firmly defined at the design stage - they can only be changed at a later date with relatively high effort." The system design of such 'deep embedded vision systems' incurs high initial costs, which can only be amortized through large quantities. As such products generally have a very low power consumption, they offer long runtimes even when operated by battery.

A current example of such a 'deep embedded vision system' is the 'RealSense' technology from Intel. These camera systems are based on the Intel 'RealSense' vision processor 'D4' with special algorithms to process the raw image streams of the integrated image sensors and calculate precise 3D depth information with high resolution at a high frame rate and output these 3D images as a result for further processing.

Another example of 'deep embedded vision systems' comes from the field of text recognition: compact modules with an integrated camera, OCR software and wireless connection are mounted directly on mechanical counters and enable automatic recording without replacing the existing counters with electronic versions. They transmit the meter readings directly to the host computer at set intervals, eliminating the need for manual readings. Due to the extremely low power consumption and the short switch-on phases, these modules offer maintenance-free running times in the region of ten years.

Flexibility through System on Chip

SoC systems enable customized image processing solutions and can be accommodated in various target systems.

© Stemmer Imaging

System-on-chip (SoC) solutions are still a young embedded computer technology with a high degree of flexibility. "SoCs allow customized systems and easy adaptation of a wide range of image sensors via standard cameras and numerous standard interfaces such as GigE Vision, USB3 Vision or MIPI. By integrating hardware such as FPGAs, GPUs or DSPs, local pre-processing and data reduction can be realized if required. Standard-compliant image distribution for further processing and standard-compliant machine communication via OPC UA are also possible," explains Keppler. When using the right software environment, such ARM-based systems under Linux also offer source code compatibility with PC systems, free programmability via C/C++ and access to image processing libraries with optimized algorithms. Keppler even believes that SoC technology has the potential to revolutionize image processing, as it only requires low initial investment and system costs and can also be easily duplicated.

Choosing the optimum system

Buzzwords such as Industry 4.0, Internet of Things (IoT) and the Industrial Internet of Things (IIoT) extension, cloud computing, distributed computing, artificial intelligence, machine learning and other technologies are an expression of the numerous developments that present users and developers of machine vision systems with major challenges when selecting the most suitable system for their respective applications. As Keppler explains, software plays a key role in finding the optimum image processing system: "This should be independent of the hardware platform as well as the operating system and, in addition, compatible with common source codes and standards in order to offer the necessary flexibility."

Keppler does not believe that classic image processing could soon become obsolete due to the rapid developments in the field of embedded vision: "Embedded vision systems have experienced an enormous upswing in recent years, both in terms of their performance and their versatility, and offer their users flexible options. However, there will always be applications in which classic PC-based image processing systems are the better solution."

Author:
Peter Stiefenhöfer is the owner of the PS Marcom Services press office in Olching.