ORB-SLAM系统特征分析研究

Analysis of system characteristics of ORB-SLAM

随着自动驾驶汽车、机器人、无人机、虚拟现实和增强现实等应用的飞速发展,其核心技术同步定位和建图(SLAM)成为目前热门研究方向之一。ORB-SLAM系统作为典型的基于特征点法的SLAM系统,具有更好的鲁棒性和更高的计算效率,无论在系统优化层面还是底层硬件架构设计层面一直被广泛关注。然而目前学术界和工业界缺乏面向ORB-SLAM系统底层硬件架构设计的系统特征分析研究。本文从跟踪线程、地图构建线程和回环检测线程出发详细介绍ORB-SLAM系统,选取了ORB-SLAM2系统进行了性能分析实验,得到了ORB特征提取和块求解器2个热点函数,并分析了2个热点函数的执行特征。在Intel i5-6500和ARM Neoverse-N1处理器平台实验对比评估了2个热点函数的IPC、分支预测失效率、一级数据缓存读失效率、最后一级缓存失效率和最后一级缓存MPKI等特征,并总结了对体系结构设计的需求,为面向ORB-SLAM系统的底层硬件架构设计提供了指导性建议。

With the rapid development of applications such as autonomous vehicles,robots,drones,virtual reality,and augmented reality,their core technology,simultaneous localization and mapping(SLAM),has become one of the current hot research fields.As a typical SLAM system based on the feature point method,ORB-SLAM has better robustness and higher computational efficiency.It has been widely concerned both at the system optimization level and the underlying hardware architecture design level.However,there is a lack of system characteristic analysis for the design of the underlying hardware architecture of the ORB-SLAM in academia and industry.This paper introduces the ORB-SLAM in detail from the tracking thread,local mapping thread and loop closing thread.The ORBSLAM2 is selected for performance analysis experiments,two hotspot functions of ORB feature extraction and block solver are obtained,and the execution characteristics of the two hotspot functions are analyzed.On the Intel i5-6500 and ARM Neoverse-N1 processor platforms,the characteristics of IPC,branch miss rate,L1D read miss rate,LLC miss rate,and LLC MPKI of the two hotspot functions are experimentally evaluated,and the different requirements for architecture design are summarized,which provides guiding suggestions for the underlying hardware architecture design of ORB-SLAM.