1.GNSS Research Center,Wuhan University,Wuhan 430079,Hubei,China
2.Electronic Information School,Wuhan University,Wuhan 430072,Hubei,China
Show less
文章历史+
Received
Published
2024-07-07
2026-02-24
Issue Date
2026-07-23
PDF (3522K)
摘要
导航系统依赖传感器感知周围环境。当前,基于单一传感器的导航系统已难以满足各类复杂场景下的导航需求,导航系统正朝传感器多源化方向发展。在多源传感器数据融合过程中,图像数据的处理最消耗时间和资源,对系统性能影响最大。为解决这些问题,设计智能导航平台的硬件控制终端,利用基于全球卫星导航系统(Global Navigation Satellite System, GNSS)秒脉冲(Pulse Per Second, PPS)的时间同步,实现多源传感器数据融合;设计用于同步定位与地图构建(Simultaneous Localization And Mapping, SLAM)前端ORB(Oriented FAST and Rotated BRIEF)特征提取加速器,加速图像处理过程,提高SLAM系统的实时性。实验结果表明,硬件平台不仅支持GNSS、惯性测量单元(Inertial Measurement Unit, IMU)、视觉和激光雷达的数据采集和融合,还能加速图像ORB特征点提取。在执行图像ORB特征提取任务时,与CPU和GPU平台上的实现相比,该加速器的帧率分别达到了它们的2.7倍和1.8倍,而功耗仅为它们的5.1%和2.9%。
Abstract
The navigation system relies on sensors to perceive the surroundings. Currently, navigation systems based on a single sensor are struggling to meet the demands of complex scenarios, leading to a shift towards multi-source sensor integration. In the process of fusing data from multiple sensors, image data processing is the most time and resource-intensive, imposing the greatest impacting on system performance. To address the issues of single-sensor navigation systems and the time and resource consumption of image processing, a hardware control terminal for an intelligent navigation platform is designed, utilizing Global Navigation Satellite System (GNSS) Pulse Per Second (PPS) for time synchronization to achieve data fusion from multiple sensors. Additionally, an ORB (Oriented FAST and Rotated BRIEF) feature extraction accelerator is designed for the frontend of Simultaneous Localization And Mapping (SLAM) to accelerate the image processing process and improve the real-time performance of the SLAM system. Experimental results show that the hardware platform not only supports data acquisition and fusion from GNSS, Inertial Measurement Units (IMU), vision, and LiDAR but also accelerates the extraction of ORB feature points. When performing ORB feature extraction tasks, the accelerator achieves frame rates 2.7 times that of the CPU implementation and 1.8 times that of the GPU implementation, while the power consumption is only 5.1% and 2.9% of theirs respectively.
现有的导航系统主要依赖全球导航卫星系统(Global Navigation Satellite System, GNSS)作为主要定位方式。为了实现更高的定位精度并满足各种复杂场景中的导航需求,这些系统通常还会配备惯性测量单元(Inertial Measurement Unit, IMU)、视觉相机和雷达等传感器,从多维度、多视角感知环境。文献[4]基于图优化算法将GNSS精密单点定位(Precise Point Positioning, PPP)和视觉相机、IMU、激光雷达4种传感器在原始观测值层面组合,显著提升了位姿估计精度、连续性和可靠性;文献[5]基于矩阵李群滤波算法实现GNSS和IMU松组合,有效提高了低精度IMU场景中的定位精度;文献[6]提出基于抗差因子图的多源信息融合定位方法,进一步降低了定位解算的平面位置误差,提升了系统的鲁棒性。以上分析表明,基于多源传感器的定位导航系统拥有更强大的环境感知能力,在复杂环境中的定位精度、可靠性都有显著提升。
多源传感器数据只在同一时间节点融合,因此,获取同一时刻的数据至关重要[7]。然而,不同传感器的时钟源各异,且各自在其时间域内运行,再加上时钟漂移的影响,即使各传感器的时间戳在初始时刻对齐,运行一段时间后,对齐的结果也会偏移。此外,不同传感器的采样频率不同,使得同步采集数据变得更加困难。因此,多源传感器系统的搭建往往离不开时间同步机制[8]。目前,多源传感器时间同步技术的主要有精确时间协议(Precision Time Protocol, PTP)、网络时间协议(Network Time Protocol, NTP)和GNSS秒脉冲(Pulse Per Second, PPS)时间同步[9]。其中,NTP是在软件层面实现同步,该技术受网络链路影响,同步精度只能达到微秒级,而PTP和GNSS秒脉冲则在硬件层面进行时间同步,两者精度均可达到纳秒级[10]。
视觉同步定位与地图构建(Visual Simultaneous Localization And Mapping, VSLAM)技术是视觉导航系统的主要组成部分之一,用于构建未知环境的地图,并利用视觉传感器在该地图中进行定位[11]。在VSLAM的解决方案中,基于特征点的SLAM对大运动和光照变化不敏感,算法占用资源也较少,相较于光流、深度学习等其他方法具有明显优势[12]。目前广泛应用的SLAM算法——ORB(Oriented FAST and Rotated BRIEF)-SLAM[13,14]利用了oriented FAST特征和rotated BRIEF(下文简写成rBRIEF)描述子。oriented FAST是在FAST特征点的基础上进行改进,它使用图像金字塔对原始图像进行处理,并用灰度质心法计算特征点的方向。BRIEF描述子是一串描述特征点的二进制向量,反映特征点邻域内像素对间的关系。rBRIEF在BRIEF基础上增加了邻域窗口的旋转模块,使其具备旋转不变性。文献[15]基于FPGA设计了图像FAST角点检测、BRIEF描述子生成和匹配的加速器架构,该架构在识别高分辨率卫星图像时的匹配率接近桌面端,但处理速度显著提升;文献[16]提出用于SLAM后端优化的列文伯格-马夸尔特(Levenberg-Marquardt, LM)算法加速器,可高效地在功耗受限的嵌入式设备上完成IMU观测信息的优化求解。因此,智能导航系统可以充分利用FPGA芯片,将计算密集的模块部署到FPGA中,提高智能导航系统的处理效率。
方向角的计算过程包括除法和反正切。Xilinx系列产品中有坐标旋转数字计算机算法(Coordinate Rotation Digital Computer, CORDIC)IP核和数字信号处理模块(Digital Signal Processing, DSP)可以精确计算反正切,但CORDIC模块对输入位宽有限制,无法计算图像矩。因此,本研究采用基于查找表(Look-Up Table, LUT)的方法近似计算方向角。首先,将圆周划分为64个扇区,每个扇区的角度为其角平分线对应的角度。其次,根据图像矩的符号确定质心所处的象限,通过取绝对值将其转换到第一象限,对应第一象限的16个角度。最后,使用4个并行的比较器比较和的值并使用优先编码器将比较的结果匹配到相应的角度。一旦确定角度,模块将象限以及输出到rBRIEF描述子生成器。
在rBRIEF描述子生成器中,采样坐标点集会旋转度,根据旋转后的点集在图像块中采样像素对。与文献[25]的方法不同,本研究采用了一种基于像素窗口的结构来执行质心计算和像素采样。在本文的rBRIEF单元中,输入的是整个像素窗口,而非传统rBRIEF单元中的单列像素数据。这种改进使得质心计算能够在并行模式下进行,同时计算窗口中行和列的像素总和。由于直接处理整个像素窗口,避免了引入额外的数据延时模块,从而简化了硬件设计并减少了潜在的处理延迟。此外,本设计不再需要4个窗口缓存存储输入列像素,而是直接在像素窗口中采样像素点。使用64个比较器并行比较像素对,经过4个时钟周期的运算,生成256比特的rBRIEF描述子。最终,模块会将角点坐标及其描述子一同输出到先入先出缓存器(First In First Out, FIFO)中,以供上位机读取。
3 实验部分
3.1 实验平台与场景
本文使用自主搭建的数据采集平台。如图12(a)所示,平台以导航小车为载体,搭载GNSS天线、机械臂、双目相机、Blackfly S USB3单目相机和Robosense 16线激光雷达。图12(b)为自主研发的硬件终端,负责导航小车的运动控制、传感器数据接收和处理。控制终端搭载u-blox F9P GNSS接收机、ADIS16470 IMU、JETSON ORIN NX计算板卡以及Xilinx Zynq XC7Z020芯片。u-blox F9P模块支持多频段实时动态定位(Real Time Kinematic, RTK)技术,具备快速收敛时间和高可靠性能,可提供厘米级定位精度。IMU采用ADIS16470模块,其性能如表1所示。JETSON ORIN NX模块是NVIDIA公司研发的用于嵌入式系统的计算板卡,它集成了8核ARM CPU和32个Tensor核心,可高效运行视觉SLAM后端计算。FPGA资源使用Zynq XC7Z020芯片的可编程逻辑(Programmable Logic, PL)端,该芯片有触发器(Flip-Flop, FF)106 400个,查找表53 200个,DSP单元220个,块内存(Block Random Access Memory, BRAM)为4.9 MB。
GUOC, LUOB H, LIF, et al. Review and verification for brain-like navigation algorithm[J]. Geomatics and Information Science of Wuhan University, 2021, 46(12): 1819-1831. DOI: 10.13203/j.whugis20210469(Ch ).
LIW B, QINC H, ZHANGT Y, et al. Review: Brain-inspired intelligent navigation modeling technology and its application[J]. Systems Engineering and Electronics, 2024, 46(11): 3844-3861 (Ch).
LIAOJ C, LIX X, FENGS Q. GVIL: Tightly-coupled GNSS PPP/Visual/INS/LiDAR slam based on graph optimization[J]. Geomatics and Information Science of Wuhan University, 2023, 48(7): 1204-1215. DOI: 10.13203/j.whugis20230119(Ch ).
GUOC, CHENY C, LUOY R. Matrix Lie group filtering algorithm for GNSS/SINS loosely integrated navigation in the world frame[J]. Journal of Chinese Inertial Technology, 2024, 32(3): 242-249. DOI: 10.13695/j.cnki.12-1222/o3.2024.03.005(Ch ).
HUANGZ R, CHAIH Z, XIANGM Z, et al. AUV multi-source information fusion localization method based on robust factor graph[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(8): 1278-1285. DOI: 10.11947/j.AGCS.2023.20210735(Ch ).
[13]
XIAK Q, XINGM, LIS M, et al. Research on time synchronization technique of multi-source composite guidance data fusion[C]//2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC). New York: IEEE Press, 2020: 1368-1373. DOI: 10.1109/ITAIC49862.2020.9338920 .
FUF J, LIC M, WUL, et al. Time calibration method for vision sensor and inertial sensor[J]. Acta Optica Sinica, 2021, 41(24):200-209. DOI:10.3788/AOS202141.2415002(Ch ).
[16]
NEELAM, RATHOREH, SHARMAL, et al. Clock Comparison with an ultra-stable optical fibre link utilizing White Rabbit Network[C]//2020 XXXIIIrd General Assembly and Scientific Symposium of the International Union of Radio Science. New York: IEEE Press, 2020: 1-4. DOI: 10.23919/ursigass49373.2020.9232324 .
WANGL Y. Design of clock synchronization platform based on ZYNQ[D]. Xi’an: Shaanxi University of Science & Technology, 2023. DOI: 10.27290/d.cnki.gxbqc.2023.000185(Ch ).
[19]
RAMOSF T, NIETOJ, DURRANT-WHYTEH F. Recognising and modelling landmarks to close loops in outdoor SLAM[C]//Proceedings 2007 IEEE International Conference on Robotics and Automation. New York: IEEE Press, 2007: 2036-2041. DOI: 10.1109/ROBOT.2007.363621 .
ZOUX, XIAOC S, WENY Q, et al. Research of feature-based and direct methods VSLAM[J]. Application Research of Computers, 2020, 37(5): 1281-1291. DOI: 10.19734/j.issn.1001-3695.2018.11.0789(Ch ).
[22]
MUR-ARTALR, MONTIELJ M M, TARDÓSJ D. ORB-SLAM: A versatile and accurate monocular SLAM system[J]. IEEE Transactions on Robotics, 2015, 31(5): 1147-1163. DOI: 10.1109/TRO.2015.2463671 .
[23]
MUR-ARTALR, TARDÓSJ D. ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras[J]. IEEE Transactions on Robotics, 2017, 33(5): 1255-1262. DOI: 10.1109/TRO.2017.2705103 .
[24]
HUANGJ J, ZHOUG Q, ZHOUX, et al. A new FPGA architecture of FAST and BRIEF algorithm for on-board corner detection and matching[J]. Sensors, 2018, 18(4): 1014. DOI: 10.3390/s18041014 .
LIUQ, LIUW Z, YUB, et al. An IMU state optimization accelerator for SLAM[J]. Journal of Beijing University of Aeronautics and Astronautics, 2023, 49(5): 1027-1035. DOI: 10.13700/j.bh.1001-5965.2021.0382(Ch ).
[27]
VEMULAPATIV, CHEND M. FSLAM: An efficient and accurate SLAM accelerator on SoC FPGAs[C]//2022 International Conference on Field-Programmable Technology (ICFPT). New York: IEEE Press, 2022: 1-9. DOI: 10.1109/ICFPT56656.2022.9974562 .
[28]
HEK M, ZHANGX Y, RENS Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916. DOI: 10.1109/TPAMI.2015.2389824 .
[29]
WEBERRUSSJ, KLEEMANL, BOLANDD, et al. FPGA acceleration of multilevel ORB feature extraction for computer vision[C]//2017 27th International Conference on Field Programmable Logic and Applications (FPL). New York: IEEE Press, 2017: 1-8. DOI: 10.23919/fpl.2017.8056856 .
[30]
ROSTENE, PORTERR, DRUMMONDT. Faster and better: A machine learning approach to corner detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2010, 32(1): 105-119. DOI: 10.1109/TPAMI.2008.275 .
[31]
SUNR D, LIUP L, WANGJ, et al. A 42 fps full-HD ORB feature extraction accelerator with reduced memory overhead[C]//2017 International Conference on Field Programmable Technology (ICFPT). New York: IEEE Press, 2017: 183-190. DOI: 10.1109/FPT.2017.8280137 .
[32]
CALONDERM, LEPETITV, STRECHAC, et al. BRIEF: Binary robust independent elementary features[C]//2010 European Conference on Computer Vision (ECCV). Berlin: Springer, 2010: 778-792. DOI: 10.1007/978-3-642-15561-1_56 .
[33]
KRAWINKELT, SCHÖNS. Benefits of receiver clock modeling in code-based GNSS navigation[J]. GPS Solutions, 2016, 20(4): 687-701. DOI: 10.1007/s10291-015-0480-2 .
[34]
TARANCOR, ARNAUJ M, GONZÁLEZA. A low-power hardware accelerator for ORB feature extraction in self-driving cars[C]//2021 IEEE 33rd International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD). New York: IEEE Press, 2021: 11-21. DOI: 10.1109/SBAC-PAD53543.2021.00013 .
[35]
PHAMT H, TRANP, LAMS K. High-throughput and area-optimized architecture for rBRIEF feature extraction[J]. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2019, 27(4): 747-756. DOI: 10.1109/TVLSI.2018.2881105 .
[36]
FANGW K, ZHANGY J, YUB, et al. FPGA-based ORB feature extraction for real-time visual SLAM[C]//2017 International Conference on Field Programmable Technology (ICFPT). New York: IEEE Press, 2017: 275-278. DOI: 10.1109/FPT.2017.8280159 .
[37]
PARKJ S, KIMH E, KIML S. A 182 mW 94.3 f/s in full HD pattern-matching based image recognition accelerator for an embedded vision system in 0.13-μm CMOS technology[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2013, 23(5): 832-845. DOI: 10.1109/TCSVT.2012.2223873 .
[38]
ZHUW P, LIUL B, JIANGG L, et al. A 135-frames/s 1080p 87.5-mW binary-descriptor-based image feature extraction accelerator[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2016, 26(8): 1532-1543. DOI: 10.1109/TCSVT.2015.2469116 .
[39]
JOHARIM M, CARTAC, FLEURETF. ESLAM: Efficient dense SLAM system based on hybrid representation of signed distance fields[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE Press, 2023: 17408-17419. DOI: 10.1109/CVPR52729.2023.01670 .