With improvements in processing power, memory density, and system integration, embedded vision is gradually expanding into emerging application areas, and its market size is expected to grow significantly within the next decade. However, as application areas become increasingly diversified, what challenges do embedded vision systems face?
An embedded vision system encompasses the entire signal chain from receiving photons from a selected image sensor to the system output. It extracts processed or unprocessed images or information from the received image and provides it to downstream systems. The embedded system architect is responsible for ensuring the performance of the receiving and output processes according to system requirements.
First, we must be familiar with the electromagnetic spectrum and the visible spectrum, which is only the wavelength range of 390 nm (blue light) to 700 nm (red light) that is visible to the naked eye. However, depending on the imaging equipment used, we can capture images of a wider range of wavelengths, including X-rays, ultraviolet light, infrared light, and the visible spectrum. For the near-infrared and lower spectral ranges, we can use charge-coupled devices (CCDs) or complementary metal-oxide-semiconductor (CMOS) image sensors (CIS); for the infrared spectrum, a dedicated infrared photodetector is required. The reason for the need for a dedicated sensor for the infrared spectrum is partly because chip image sensors such as CCDs or CISs require excitation energy. These components typically require 1 eV of photon energy to excite an electron; however, in the infrared spectrum, photon energy is between 1.7 eV and 1.24 meV, therefore infrared image sensors should be based on HgCdTe or InSb. These low excitation energy sensors are often used in conjunction with CMOS readout ICs (ROICs) to facilitate sensor control and readout.
The two most common sensor technologies are CCD and CIS:
Charge-coupled devices (CCDs) are considered the best analog components; therefore, integration with digital systems requires the addition of an analog-to-digital converter (ADC) and frequency generation functionality at a given analog voltage. Each pixel stores the charge generated by photons, and most applications use a 2D array arranged in rows, with each row consisting of multiple pixels. Reading a CCD involves transferring each row in parallel to a read buffer, which then reads each row serially. During this reading process, the charge is converted into voltage.
CMOS image sensors allow for tighter integration of ADCs, bias circuits, and drive circuits on a single chip, significantly reducing system integration requirements while simultaneously increasing the complexity of CIS design. Active pixel sensors (APS) are the core of CIS, differing from CCDs in that each pixel in a CIS contains a set of photodiodes and a readout amplifier, enabling the independent reading of any single pixel in the array. Although most embedded vision systems utilize CIS components, CCDs remain the primary sensor used in high-end scientific research applications.
CMOS and CCD are currently the two leading technologies used in image acquisition. CCD offers higher image quality, but over the past decade, the gap between CMOS and CCD has narrowed significantly, with CMOS showing a strong trend towards surpassing CCD in terms of power consumption and cost. Furthermore, many applications require efficient parallel processing systems, necessitating dedicated hardware processors such as GPUs, DSPs, FPGAs, and multi-core SoCs. However, this undoubtedly increases system cost, power consumption, and PCB size. Therefore, a cost-effective processor is also essential in the industry. In practical applications, we must select a suitable processor based on the system's real-time performance, power consumption, image accuracy, and algorithm complexity. To assist users in building their own embedded vision platforms and products, Avnet offers a series of vision application solutions, such as the PicoZed Embedded Vision Development Kit. The PicoZedSoM integrates a Xilinx Zynq-7030 AllProgammable SoC, and also includes the PicoZed Expansion Board V2.0, an HDMI FMC Expansion Board (with an integrated camera interface), and a Python-1300-CSXGA (1280x1024) camera module. This PicoZed embedded vision development kit is suitable for developing advanced vision applications. In addition to hardware, software tools, and a wealth of licensed IP resources, it also supports the reVISIONStack technology stack. reVISIONStack includes abundant design resources such as algorithms, hardware-accelerated OpenCV libraries, and currently popular neural network training datasets. Embedded vision systems are constantly evolving, and with the efforts of major manufacturers and engineers, various bottlenecks will be overcome, leading to wider applications in fields such as cluster vision, artificial intelligence, the Internet of Things, and industrial automation.
Facing bottlenecks and challenges is not terrible; what is terrible is standing still and not moving forward. With the development of technology, embedded vision systems are becoming increasingly diversified.
