LAUSR

Binocular stereo vision is a commonly applied computer vision technique with a wide range of applications in 3D scene perception. However, binocular stereo matching algorithms are computationally intensive and complicated. In addition, some traditional platforms are unable to meet the real-time and energy efficient dual requirements. In this paper, we proposed a hardware/software co-design FPGA (Field Programmable Gate Array) approach to overcome these limitations. Based on the characteristics of binocular stereo vision, we modularize the system functions to achieve the hardware/software partitioning. This accelerates the data processing on the FPGA, while simultaneously performing data control on the ARM (Advanced RISC Machine) cores. The parallelism of the FPGA allows for a full-pipeline design that is synchronized with an identical system clock for the simultaneous running of multiple stereo processing components, thus improving the processing speed. Furthermore, to minimize hardware costs, the collected images and data are compressed prior to matching, while the precision is subsequently enhanced during post-processing. The proposed system was evaluated on the PYNQ-Z2 development board, with experimental results revealing its high real-time performance and low power consumption for a 100M clock frequency. Compared with existing designs, the simple yet flexible system demonstrated a higher image processing speed and less hardware resource overhead (thus lower power consumption). The average error rate of the BM matching algorithm was also improved, particularly with the limited PYNQ-Z2 hardware resource. The proposed system has been opened on GitHub.