Development of a stereo vision measurement architecture for an underwater robot

Underwater robotics tasks are considered very critical, mainly because of the hazardous environment. The embedded systems for this kind of robots should be robust and fault-tolerant. This paper describes the development of a system for embedded stereo vision in real-time, using a hardware/software co-design approach. The system is capable to detect an object and measure the distance between the object and the cameras. The platform uses two CMOS cameras, a development board with a low-cost FPGA, and a display for visualizing images. Each camera provides a pixel-clock, which are used to synchronize the processing architectures inside the FPGA. For each camera a hardware architecture has been implemented for detecting objects, using a background subtraction algorithm. Whenever an object is detected, its center of mass is calculated in both images, using another hardware architecture to do that. The coordinates of the object center in each image are sent to a soft-processor, which computes the disparity and determines the distance from the object to the cameras. A calibration procedure gives to the soft-processor the capability of computing both disparities and distances. The synthesis tool used (Altera Quartus II) estimates that the system consumes 115.25mW and achieves a throughput of 26.56 frames per second (800×480 pixels). These synthesis and the operation results have shown that the implemented system is useful to real-time distance measurements achieving a good precision and an adequate throughput, being suitable for real-time critical operation.