基于Pynq的半全局立体匹配算法实现

Implementation of a Semi-global Stereo Matching Algorithm Based on Pynq

立体匹配算法作为双目立体视觉技术的核心算法,具有运算复杂度高、资源占用率高、实时性差的特点。而现有的基于FPGA实现方法多采用高性能FPGA芯片,利用其板卡丰富的资源进行纯verilog语言开发。这样做虽然实时性效果不错,但存在开发难度较大、周期较长、成本较高的弊端。论文针对该问题,提出了一种基于相对低廉的Pynq-z2开发板在PL+PS端协同开发的SGM立体匹配算法,将运算量大的部分在FPGA端进行硬件加速,部分代价优化算法在ARM端利用Python语言实现。实验表明,该算法可有效降低资源占用率,基本满足在相对低廉的嵌入式设备中实现的要求。

As the core algorithm of binocular stereo vision technology,the stereo matching algorithm has the characteristics of high complexity,high resource consumption and poor real-time performance.The existing FPGA-based implementation methods mostly use high-performance FPGA chips,its board rich resources are used to develop pure verilog language.Although the re⁃al-time effect is good,but there is a development difficulty,long cycle time,high cost of the disadvantages.In this paper,an SGM stereo matching algorithm is proposed based on the relatively inexpensive Pynq-z2 development board developed collaboratively on the PL+PS side,where the computationally intensive part is hardware accelerated on the FPGA side and part of the cost-optimised algorithm is implemented in the ARM side using the Python language.Experiments show that the algorithm can effectively reduce the resource occupancy rate and basically meet the requirements of relatively inexpensive embedded devices.