计算机工程

• •    

改进的实时半全局立体匹配算法及硬件实现

  

  • 发布日期:2020-12-08

Hardware Implementation of an Improved Real-Time Semi-Global Stereo Matching Algorithm

  • Published:2020-12-08

摘要: 为提高传统的基于 FPGA 的实时立体匹配系统的匹配精度,改善传统 Census 变换算法在特定区域误 匹配率较高的缺点,提出一种具有高并行性流水线结构的实时半全局立体匹配算法。以改进的 Tanimoto 距 离和带权重 4 方向的梯度绝对值差(ADG)的组合作为新的初始匹配代价,代价聚合采用 4 路径并行结构 的 SGM 算法,视差选择采用赢家通吃策略(winner-takes-all,WTA),视差校正过程采用阈值检测(threshold detection,TD)算法代替传统左右一致性检验算法。结果表明,改进的立体匹配算法提高了弱纹理和边缘 区域 的 区分度,减少了对中心点的依赖,降低了资源占用;在 Middleburry 测试平台上,算法平均误匹配率 为 7.52%;在 Xilinx Zynq-7000 平台上时钟频率为 100MHz、最大视差为 64 时,算法的吞吐率为 450×375/98FPS。研究对于提高实时立体匹配技术水平具有现实意义。

Abstract: In order to increase the matching accuracy of traditional real-time stereo matching algorithm based on FPGA and solve the problem on high mismatching rate in specific regions calculated by traditional Census transform algorithm, a real-time semi-global stereo matching algorithm with high parallelism and pipeline architecture was proposed. The combination of the improved Tanimoto distance and the weighted 4-direction absolute differences of gradient (ADG) method were set as the initial matching cost, and a four-path semi-global matching algorithm was used in the cost aggregation stage. The winner-takes-all strategy was chosen in the disparity computation, and the truncation threshold detection method replaced the traditional left-right check to reduce the resource usage in the post-processing step. The experimental results demonstrate that the proposed algorithm improves the discrimination in the weak texture and edge region, with the less dependence on central point and less resources usage. The algorithm achieved an average error rate at 7.52% when evaluated on the Middleburry benchmark, and the hardware architecture implemented on XilinxZynq-7000 FPGA resulted in a throughput of 450×375/98FPS with 64 disparity levels at 100MHz. The research has the practical significance for improving the level of real-time stereo matching technology. Key words FPGA; stereo matching; Census Transform; Tanimoto distance; real-time semi-global matching;