Digital video has become a killer app for signal processing technologies, and video analytics—that is, analysis of digital video to identify specific events or characteristics—is quickly becoming a significant driver in digital video. Video analytics isn’t one of those solutions looking for a problem; it has an enormous range of potential applications, both commercial (such as intelligent surveillance and traffic monitoring) and military (such as target detection and tracking).
Like most video applications, however, video analytics is computationally demanding and bandwidth-intensive. It’s hard to do it well; the algorithms tend to be extremely complicated. Furthermore, the analysis often has to be done in real-time, using streaming data. As a result, the challenge in creating an effective video analytics solution is threefold: you need clever, robust algorithms to identify the desired events/characteristics; you need a processing engine that’s powerful enough to run those clever algorithms in real-time; and you need an overall system architecture that won’t choke on the data. Newcomer Eutecus claims to have all three.
Eutecus, based in Berkeley, California, has developed a set of proprietary analytics algorithms, and taken them into hardware by creating a specialized FPGA-based multi-core processing architecture. The processing engine was designed to run the algorithms efficiently, and conversely, the algorithms were designed to enable efficient hardware-based implementations.
The resulting hardware/software combination is intended to be located right next to the camera sensor, thus eliminating the need to constantly transmit the full video stream across, say, a corporate IP network. Instead, the in-camera analytics processor can potentially just transmit an “alarm” or partial video stream when there’s a designated event, which requires considerably less bandwidth. Multiple cameras can be networked together, and their partial video streams can be further processed (at a server) for higher-level analysis—such as tracking a person who has moved between the areas monitored by two cameras. Eutecus’ localized approach also facilitates adaptive and high-dynamic range sensing by allowing real-time adjustment of parameters such as integration time and sensor gain.
Eutecus is comprised of video algorithm and processor architecture experts, several of whom—including co-CTOs Csaba Rekeczky and Akos Zarandy—are originally from Hungary. Their analytics algorithms are based on what Eutecus calls “Cellular Vision Technology,” a massively parallel processing approach that’s well suited to highly parallel hardware implementation.
Thus far, the company’s funding has been in the form of government grants (SBIR and STTR) and its initial product (a custom ASIC running video analytics algorithms) is geared for military use—but with the FPGA-based approach, Eutecus is now moving into the commercial realm. Eutecus has partnered with Xilinx for the project, and the initial implementations of its software and multi-core processing engine are on Spartan-3 chips (Eutecus is in the process of migrating its technology to the Spartan-3A DSP family).
The FPGA-based processing engine is dubbed “C-MVA,” for Cellular Multicore Video Analytics, and a single C-MVA can include from 64-128 cores. The C-MVA video analytics engine will typically be coupled to one or more backend processors that run an OS and handle control and communications.
Exactly what’s on the FPGA can be tailored for customer’s needs; it can have multiple C-MVA cores, up to two microprocessors, various types of I/O (such as a UART), an SDRAM controller, etc. The idea is to create a complete video analytics SoC. The company says that current implementations (in a Spartan-3 XC3S4000) can process the lower end of HD resolution (720x1024) at 30 fps in 1 Watt.
Eutecus isn’t selling chips, however; what they’re selling is reference designs for FPGA-based video analytics SoCs plus the tools to create a custom design—and then charging a license fee plus royalties for use of their proprietary algorithms and hardware design. According to Eutecus, the incremental cost of its solution for a network of about 100 surveillance cameras will be in the range of several hundred dollars per camera, which it claims is an order of magnitude less than server-based solutions.
Add new comment