基于无监督学习的ToF点云目标提取方法

飞行时间(Time of Flight,ToF)三维成像技术在人工智能领域具有重要的应用价值。间接ToF三维成像是通过向目标发射调制的光强信号,再经过目标反射到相位解调图像传感器获得相位差,通过计算获得目标的深度信息。由于间接ToF成像技术会受到背景界面多次反射产生的多路径干扰,因此在复杂环境中目标物体深度测量数据会受到侧面和背景界面的多次反射的回波信号影响,降低边缘处深度测量的精度水平,因此需要对原始点云数据进行目标提取和多路径去除的预处理。本文针对该问题提出一种多界面场景中基于点云矢量的目标提取方法,能够实现复杂多目标的快速提取和多路径强干扰的去除。首先基于kmeans提出一种FVPkmeans算法,完成目标点云数据的全局全矢量提取处理。再基于K NN提出一种迭代滤波算法,实现局部多路径干扰数据的滤除。通过与其它方法的比较研究,该方法能够有效去除TOF点云目标数据的多路径干扰,目标提取性能提高了40,实验表明本文提出的全局点云数据全矢量目标提取和多路径干扰去除算法能够实现对目标点云数据的无监督学习智能提取与滤波要求。

ToF point cloud target extraction method based on unsupervised learning

Time of flight(TOF)3D imaging technology has important applications in the field of artificial intelligence.By transmitting a modulated light intensity signal to the target,then reflecting it to the phase demodulation image sensor to obtain the phase difference,the depth information of the target,indirect ToF 3D imaging can be obtained through calculation.Because indirect TOF imaging technology is subject to multi path interference caused by multiple reflections from the background interface,multipath interference(MpI)affects the ToF depth measurements.Therefore,it is necessary to preprocess the original point cloud data for target extraction and multi path removal.To solve the above problem,a target extraction method based on point cloud vector in multi interface scene is proposed in this paper.The method can achieve the rapid extraction of complex multi target and the filtering of MpI.Firstly,a FVP k means algorithm based on k means is proposed to complete the global full vector extraction of target point cloud data.Then,an iterative filtering algorithm based on K NN is proposed is proposed to realize the filtering of local multi path interference data.Compared with other methods,the method can effectively remove multi path interference from TOF point cloud target data and the target extraction performance is improved by 40%.The experiments show that the global point cloud data full vector target extraction and multi path interference removal algorithm proposed in this paper can realize the unsupervised learning intelligent extraction and filtering requirements for target point cloud data.