Efficient Hardware Platform Architecture for High-Precision Stereo Vision Processing

To achieve high-precision and efficient endoscopic measurement, this study introduces a highly parallel hardware architecture dedicated to the implementation of binocular stereo matching algorithms, aiming to enhance real-time performance while ensuring accuracy. The system employs a combination of the CENSUS and local-expansion (LE) algorithms for precise disparity extraction. Through pipeline design and optimized memory partitioning, we have enabled efficient parallel operation of various system modules. Experimental results demonstrate that the pipeline implementation increases the processing speed by $\mathbf{1 0 7. 3 \%}$ compared to the serial implementation. On the medical dataset SERV-CT, our algorithm exhibits the highest accuracy. Furthermore, the universality of this hardware architecture allows it to adapt to the implementation of various traditional algorithms, thereby enhancing the system’s applicability and scalability.