Abstract
The RazorCam project introduces a cooperative smart camera system designed as a flexible and extensible hardware/software environment for prototyping and verifying video applications. The platform utilizes a Xilinx Spartan-6 LX FPGA (Trenz Electronic GigaBee XC6SLX micromodule) as the main computing host, integrating a Gigabit Ethernet transceiver and two independent banks of 128 MB DDR3 SDRAM. The system supports modular camera heads, including a Tau 640 infrared sensor from FLIR and an MT9V034 wide VGA CMOS digital sensor from Aptina. Output is handled via a 320 x 240 TFT LCD with a serial peripheral interface. By porting Intel's OpenCV library to an embedded Linux environment running on a MicroBlaze soft processor, RazorCam enables the seamless use of hardware accelerators through a dedicated streaming data interface, overcoming traditional bus bottlenecks in real-time video processing.
Introduction
Computer vision has traditionally been the domain of desktop processors and GPUs. However, as vision applications move toward mobile and embedded environments—such as drones, smart surveillance, and driver assistance—developers face a "performance wall." Standard embedded processors often lack the raw power to handle high-definition video streams in real-time. While FPGAs (Field Programmable Gate Arrays) offer the necessary speed through hardware parallelism, they are notoriously difficult to program, requiring low-level hardware description languages (HDL). RazorCam addresses the need for a platform that maintains the ease of software development while leveraging the massive speedups of FPGA hardware acceleration.
System Architecture
The core of the RazorCam platform is a Xilinx Spartan-6 FPGA (XC6SLX45) acting as the central hub. It hosts a MicroBlaze "soft" processor core running an embedded Linux OS, providing a solid, familiar platform for development with a feature-rich toolchain.
In traditional architectures, the system bus often becomes a bottleneck when moving large amounts of pixel data. RazorCam solves this by implementing a dedicated streaming data interface for hardware accelerators, ensuring data flows seamlessly without clogging the CPU's main AXI bus. This modular architecture allows for various "camera heads" to be swapped depending on the application requirements, such as infrared or digital imaging.
Design Methodology
To simplify development, the authors propose a methodology that moves from abstract code to physical hardware through a structured four-step design flow:
- System Specification: Define the application behavior using standard C++ and OpenCV.
- HW/SW Partitioning: Decide which parts stay in software and which move to hardware. Hardware parts are modeled using SystemC and Transaction Level Modeling (TLM) for early verification.
- RTL Implementation: Translate the SystemC descriptions into VHDL or Verilog for the FPGA logic.
- Emulation: Run the final system on the RazorCam hardware under Linux.
Performance Evaluation
The platform was evaluated using a standard image convolution task, a common but computationally expensive operation in computer vision. The benchmarks compared a software-only implementation against the hardware-accelerated version on the RazorCam hardware:
- Software-only (OpenCV on MicroBlaze): Achieved less than 1 frame per second (fps).
- Hardware-accelerated (FPGA Logic): Achieved 330 fps for VGA (640x480) resolution.
- Efficiency: This 330x speedup was achieved while running at a modest 100 MHz. Furthermore, moving the task to hardware only increased Look-Up Table (LUT) usage from 31% to 33%, demonstrating high resource efficiency.
Conclusion
RazorCam successfully bridges the gap between high-level computer vision software and high-performance hardware acceleration. By providing a modular environment and a structured design flow, it enables researchers to prototype real-time image processing applications that are significantly faster than software-only embedded implementations. Future work will focus on automated design space exploration for task partitioning and testing the platform in complex real-world scenarios like airport surveillance and in-vehicle driver assistance systems.
Summary updated