刘强^{1, 2}，秦书臻^{1, 2}，俞波³，刘少山³
AuthorsHTML:刘强^{1, 2}，秦书臻^{1, 2}，俞波³，刘少山³
AuthorsListE:Liu Qiang^{1, 2}，Qin Shuzhen^{1, 2}，Yu Bo³，Liu Shaoshan³
AuthorsHTMLE:Liu Qiang^{1, 2}，Qin Shuzhen^{1, 2}，Yu Bo³，Liu Shaoshan³
Unit:1. 天津大学微电子学院，天津 300072；
2. 天津市成像与感知微电子技术重点实验室，天津 300072；
3. 深圳普思英察科技有限公司，深圳 518000
Unit_EngLish:1. School of Microelectronics，Tianjin University，Tianjin 300072，China；
2. Tianjin Key Laboratory of Imaging and Sensing Microelectronic Technology，Tianjin 300072，China；
3. Shenzhen PerceptIn Technology Co.，Ltd.，Shenzhen 518000，China
Abstract_Chinese:光束平差法是一种广泛应用于视觉SLAM系统后端的优化算法，但在实际应用中一直存在实时性差、能耗高的问题．本文在嵌入式FPGA平台上实现了一款针对相机位姿已知的光束平差算法的加速器．该加速器为计算中最耗时的雅可比矩阵更新和成本函数计算模块定制化硬件计算架构．首先分析算法结构去除冗余计算，再根据算法数据相关性设计流水线和并行处理单元，利用时分复用的方法将雅可比矩阵更新和成本函数计算在同一硬件上实现，同时在算法中引入猜测执行进一步提高计算速度，既满足了性能要求，又减小了硬件消耗．在Xilinx xc7z020上的实验结果表明，本文设计的光束平差法加速器，相对于嵌入式ARM处理器有5.9倍的性能提升，同时节约能耗68.7%．
Abstract_English:Bundle adjustment(BA)is an optimization algorithm widely used in the backend of visual SLAM systems，but it exhibits the problems of low real-time performance and high energy consumption in practice. This paper designs an embedded FPGA accelerator for BA with a known camera pose. The accelerator customizes hardware computing architectures for Jacobian matrix update and cost function calculation which are the most time-consuming parts of the BA. The bundle adjustment algorithm is first analyzed to remove redundant calculations，and then the pipelined and parallel processing unit is designed based on the algorithm’s data dependence. Using time division multiplexing，Jacobian matrix update and cost function calculation are implemented on the same hardware，and the speculative execution is incorporated into the algorithm to further improve computation speed. By doing so，the hardware resources are saved and the performance requirements are satisfied. The experimental results on Xilinx xc7z020 indicate that the accelerator has a 5.9 fold performance improvement，compared with the embedded ARM processor. At the same time，the accelerator can save 68.7% of energy.
Keyword_Chinese:光束平差法；现场可编程门阵列；同步定位与地图构建
Keywords_English:bundle adjustment；field programmable gate array；simultaneous localization and mapping

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6557